rcekit 2.35.1__tar.gz → 3.0.0__tar.gz

rcekit-3.0.0/CHANGELOG.md

@@ -0,0 +1,2747 @@
+# Changelog
+
+All notable changes to RCEKit are recorded here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and RCEKit follows
+[Semantic Versioning](https://semver.org/spec/v2.0.0.html): PATCH for fixes,
+MINOR for new capabilities, MAJOR for breaking changes to the CLI, output
+formats, or the template schema.
+
+## [Unreleased]
+
+### Added
+
+- **A coverage ledger, so a measured result and a reproduced one stop looking
+  alike.** The README's CVE table had four rows and two columns that mattered,
+  and no way to say what stood behind any of them. It gained seven columns and
+  five more rows, and the two on the right are the point: `Bench case` says
+  whether `tests/bench/` reproduces the row under Docker with its negative
+  control, and `Recording` says whether a file in `confirmation-gifs/` shows it.
+
+  Five rows are new, measured against vulhub builds during this work: OpenTSDB
+  2.4.1 (CVE-2023-25826) reaching `confirmed` through `oob` and `needs-review`
+  through `time`; HugeGraph 1.2.0 reaching `confirmed` through both `eval` and
+  `reflected`; and a Spring Boot application on fastjson 1.2.83 reaching
+  `deserialization-sink`. All five landed saying `not yet` under `Bench case`,
+  which was the honest state at the time and the reason that column exists; each
+  has a case now. The 4 rows that predate this work had last been executed as a
+  set at 2.36.0, and none of them regressed. The ledger carries the figures for
+  the whole set, because a running total written here rots on the next change
+  rather than the next release.
+
+  `Advisory` is empty on three of them, deliberately. HugeGraph 1.3.0 answers
+  the arithmetic exactly as 1.2.0 does -- its Gremlin API evaluates Groovy
+  unauthenticated by design -- and fastjson resolving an `Inet4Address` is
+  documented autoType behaviour. Those rows prove a *method* against real
+  software, which is worth recording; calling them CVE reproductions would be
+  the overclaim the table exists to avoid. Both facts were measured rather than
+  assumed: the patched HugeGraph image was pulled and run.
+
+  Three tests hold the ledger to the repository: a row claiming a bench case
+  must have one, a row claiming a recording must have one, and the capability
+  rows must stay distinguishable from the CVE rows. Each was checked by making
+  the false claim and watching it fail. The tier check now reads *every* row
+  rather than only those naming an advisory, and its parser is keyed on column
+  headings rather than position -- the old one was hard-coded to four columns
+  and found nothing at all once the table grew.
+
+- **A bench case for `write`, the last tier-1 method without one.**
+  `tomcat-cve-2017-12615` -- `PUT /rcekit-probe.jsp/` with the content in the
+  body, which is the shape the method's own docstring names. Its control runs
+  `reflected` against the same point and must stay `negative`: the target is
+  exploitable and `write` confirms on it, but a PUT answers 204 with an empty
+  body, so nothing in the vulnerable response is computed. Measured at 959
+  probes, every one answered 204.
+
+  It is the first case here that may not share a target between its halves --
+  both write a file into the web root -- which turned up a harness bug: under
+  `--keep-up` the teardown *between* the halves was skipped too, and compose
+  handed the control the container the vulnerable half had just written to. That
+  teardown is mandatory now wherever the halves would land on the same
+  container, which is decided from what *starts* a target -- its `vulhub_path`
+  and `compose` -- and not from what stops it. A control overriding only
+  `compose_down`, to keep its volumes say, still meets the container the
+  vulnerable half brought up. A control that brings up its own target has
+  nothing to contaminate, so `--keep-up` leaves the vulnerable one running
+  there, which is the environment worth looking at after a failure. The
+  teardown after the last half stays the flag's to skip either way.
+
+  The case is also what caught the two `write` defects fixed in 2.45.6 and
+  2.45.7, neither of which a fixture could have found.
+
+### Fixed
+
+- **The comparison table named seven of the nine registered methods.** `lookup`
+  and `boolean` had no row in "How RCEKit compares", although `lookup` is the
+  Log4Shell row the CVE table directly above it already carries, and `boolean`
+  shipped two releases ago. The table's claim is coverage -- "every class, one
+  run" -- so a method missing from it is that claim understating itself.
+
+  The drift was structural, not an oversight anyone could have caught by
+  reading. `docs/reference.md`, the `--methods` help, the CVE rows and the demo
+  headings are each pinned to `DETECTION_METHODS` by a test. The comparison
+  table was pinned by nothing, so it was the one place a new method could fail
+  to appear without anything going red.
+
+  Two tests now hold it, in both directions: every registered method has a row,
+  and the table names no method the CLI would reject -- because a one-way check
+  rots. A method *removed* from the registry would otherwise leave its row
+  behind, advertising a `--methods` value that no longer exists, while the
+  completeness test went on passing.
+
+  Each row now names its method in backticks, which is what makes the check
+  possible and also answers the question the table used to leave hanging: a
+  reader who wants the row they just read has the flag in front of them. Rows
+  that are not methods name none on purpose -- second-order execution, the
+  query-language bridges and the per-dialect Windows probes are things the
+  methods run *through*, not entries in `--methods`.
+
+  Both checks read the capability table alone, and identify a method label
+  structurally: the last backtick in the leading cell, after an em-dash. The
+  first attempt did neither, and both shortcuts cost it the guarantee it stated.
+  Reading every table in the section meant a deleted capability row kept passing
+  on a mention elsewhere -- the engagement-controls table names `file` and
+  `write` under "Cleanup commands" -- which was confirmed by deleting the `file`
+  row and watching the check stay green. And filtering labels on lowercase
+  letters exempted exactly the spellings a typo or a rename produces:
+  `boolean2`, `second-order`, `reflected_v2` all walked past the phantom check,
+  while `cmd.exe` is the legitimate backtick in that column that the filter had
+  been reaching for. Structure separates them; a character class does not.
+
+  No version bump -- this is documentation and tests only.
+
+### Added
+
+- **A tier correction has to reach the prose, not only the class.** Four tests
+  now read the tier from `DETECTION_METHODS[name].tier` and hold the sentences
+  an operator actually sees against it.
+
+  `lookup` moved from `confirmed` to `lookup-sink` on the class, and the move
+  was made in the README's CVE table -- but `blind_sink_advice` went on
+  offering `--methods lookup` as a method that "confirms", in the list where
+  `oob` and `file` do mean confirmed execution, and the README's Log4Shell demo
+  heading still said `confirmed` over alt text calling the run
+  "auto-confirming a blind Log4Shell RCE". Three places, one correction, and
+  nothing compared them: the existing test pinned the one line it was written
+  for (`time` is marked needs-review only) rather than asking every line the
+  same question.
+
+  So: every `blind_sink_advice` line naming a method must state that method's
+  tier and no other; every CVE row must state the tier its method reports;
+  every demo heading must match its row in that table; and a recording below
+  `confirmed` may not be described as confirming, alt text included, since
+  that is the sentence a screen reader reads out.
+
+  The wording checks match stems rather than words, and denials such as
+  "never confirmed" are removed before they are applied. "confirms" alone
+  would have readmitted the overclaim through "confirming RCE" or "confirmed
+  execution" -- an assertion answering the same way for the right reason and
+  the broken one, which is the defect being guarded against. Each test was
+  checked by reintroducing the defect it exists for, and each carries a floor
+  on how much it inspected: a parser that quietly stopped matching would
+  otherwise pass exactly as a clean README does.
+
+  Test-only: no version bump, and nothing about a run changes.
+
+- **The same check, for the two tables under `docs/`.** They were left out when
+  the README's were pinned, and they carry the same claim:
+  `docs/reference.md` names the tier each method can reach, and
+  `docs/guide.md` tells an operator which method to reach for next.
+
+  `reference.md` states a ceiling, so its rule is a subset rather than an
+  equality -- a cell may also name a weaker tier the method really emits, as
+  `write` and `deser` both do -- and every registered method must have a row,
+  so a capability cannot land unlookupable. `guide.md` has no tier column and
+  so is held only to the negative: a row recommending a method that cannot
+  confirm may not describe confirmation.
+
+  Rows are read by column heading rather than by position. Reading the whole
+  row made the `guide.md` check skip the one row it was written for, because
+  that row's prose names `oob` -- which confirms -- while recommending
+  `lookup`, which does not.
+
+  A reference row naming something that is not a registered method fails rather
+  than being passed over. Skipping it left the completeness check one-way: a
+  method *removed* from `DETECTION_METHODS` would leave its row behind, every
+  remaining row would still match, and the page would go on offering a
+  `--methods` value the CLI rejects.
+
+  Rows split on *unescaped* pipes. Markdown writes a literal pipe in a cell as
+  `\|`, which `docs/reference.md` already does in three tables, and splitting
+  on every pipe invents a cell: zipping against the header then drops the last
+  column outright, so a claim there stops being examined and every one of these
+  checks passes without looking at it.
+
+  A disclaimer is not read as a claim. "never confirmed" was already removed
+  before the `confirm` stem was looked for; "without confirmation" and
+  "unconfirmed" were not, and the second carries its negation inside the word
+  where a rule about preceding words cannot see it. Documentation saying the
+  honest thing would have failed the suite. One parser and one denial rule now,
+  where there were two of each here and a third in the generator tests.
+
+### Changed
+
+- **A bench case may bring its target up once for both halves**, with
+  `"share_target": true`. Bringing the container up twice is the largest fixed
+  cost in a case and both halves usually hit the same one, so on a fast case it
+  is most of the wall clock.
+
+  It is opt-in, and the default is unchanged, because the teardown between the
+  halves is `down -v`: today's control meets a **fresh** target. A case whose
+  vulnerable half writes a file, plants a shell or changes a setting would hand
+  its control a target it had already altered, and a control measured against a
+  contaminated target measures nothing -- which is the one failure a benchmark
+  may not have. Validation rejects the key on a case whose control brings up a
+  different target, since there is then nothing to share and leaving it set
+  would read as though there were.
+
+  A shared `up` that fails falls back to per-half management rather than
+  carrying on, so the run reports `compose up failed` instead of two readiness
+  timeouts naming the wrong cause.
+
+  Both shipped cases set it, because neither half writes anything. Measured on
+  `struts2-s2-001` against vulhub on Docker: 33.8s to 22.9s, with both halves
+  reaching the same verdicts either way.
+
+- **The benchmark reaches a callback method**, with a case for Log4Shell
+  (Apache Solr 8.11.0, Log4j 2.14.1). `python tests/bench/runner.py --all` is
+  3/3.
+
+  Every case until now was in-band, so the listener, the token correlation and
+  `confirm_each` had never run against real software -- the area that produced
+  three P1 findings while `lookup` was being written.
+
+  Two things had to give. A JNDI lookup resolves through the system resolver,
+  which asks UDP 53, and a developer machine rarely has that port free; a bare
+  `--oob-host` IP is no way around it, because a lookup has no second channel
+  to carry the token. So a case may now name `run_in`, and the run happens in a
+  container on the target's own network at a fixed address, with an override
+  pointing the service's `dns:` at it. The repository is mounted read-only and
+  the image is a stock Python.
+
+  The other was the harness's own vocabulary. `VALID_EXPECTATIONS` and
+  `CONTROL_EXPECTATIONS` were written out by hand and had drifted: neither
+  `lookup-sink` nor `deserialization-sink` was in either, so a case for `lookup`
+  or `deser` could not be *loaded*, let alone run. Both are read from
+  `DETECTION_METHODS` now.
+
+  The control is the case: `oob` against the same Solr comes back `negative`
+  **although it is exploitable**, because every probe it builds is a shell
+  command and a `${jndi:...}` sink runs none of them. That is the gap `lookup`
+  was added to close, and until this case ran it rested on a fixture.
+
+  `{bench}` / `{repo}` now expand in a case's compose argv as well as its
+  invocation, and resolve to the mount point when the run is containerised.
+
+  The results file is created before the container starts, and it rather than
+  its directory is made writable. Under Docker's user-namespace remapping,
+  container root is a subordinate host UID, so a `mkdtemp` owned by the runner
+  at 0700 is not writable from inside: the file would never appear and the case
+  would report `nothing-tested`, as though detection had found nothing rather
+  than as though the channel had been shut. Granting the one file and not the
+  directory keeps anyone else from creating, replacing or unlinking entries
+  there.
+
+  Test-only: no version bump, and nothing about a run changes.
+
+### Fixed
+
+- **`--methods` enumerated only some of the methods it accepts.** `write`,
+  `lookup` and `deser` were registered in `DETECTION_METHODS` and had never
+  once been named in the help, so `--help` described a whole target class as
+  out of reach -- an upload that stores a file, a `${jndi:...}` sink, an
+  endpoint that deserializes what it is handed -- while the method for it was
+  already shipping and documented everywhere else. Each now carries the tier
+  its class declares: `confirmed` for `write`, with `needs-review` for a write
+  that is served but not interpreted; `lookup-sink` for `lookup`;
+  `deserialization-sink` for `deser`; and neither of the last two ever
+  `confirmed`.
+
+  `oob` also called itself "the only confirmed-tier method for a fully blind
+  sink". The tier is right, but an exclusivity claim is the kind that goes
+  stale without anything failing, so the line now says what the class itself
+  says: it confirms a sink with no output channel and no writable web root.
+
+  A new test in `CLIDocumentationTestCase` holds the list to the registry, the
+  way `--eval-engines` is already held to the corpus. It reads the `--methods`
+  help alone rather than the whole page, because every one of these names also
+  occurs inside some other flag -- `file` in `--request-file`, `write` in
+  `--file-write-path`, `time` in `--time-base` -- and it looks for the name
+  followed by the parenthesis that opens its description, because the old help
+  carried the word `write` inside `file`'s "write+read-back" and a looser
+  match would have counted that as documentation.
+
+  A second pass corrected what `--oob-host` promises those two methods. Its
+  help offered "an IP it can reach or a domain delegated to this listener" and
+  was scoped `(--methods oob)`, which was true when `oob` was the only caller.
+  It is not true now: `LookupCallback.build_probes` returns no probes at all
+  for an address literal, so `lookup` reports `nothing-tested`, and `DeserSink`
+  builds `<token>.<IP>`, which resolves nowhere, so its DNS gadget can never
+  reach `deserialization-sink`. An operator who satisfied the stated
+  prerequisite got silence from one method and a capped tier from the other.
+  The flag's help, the `--methods` entries for `lookup` and `deser`, the
+  message printed when a callback method is selected without a host, and the
+  `deser` oracle table in `docs/reference.md` now all say the same thing: an IP
+  serves `oob` alone, because only `oob` can put its token in a URL path.
+
+  `boolean` landed on `main` while this branch was open. It was named in the
+  help but left out of the `--methods` row in `docs/reference.md` -- the same
+  drift, one table over. Both carry it now, and the new test covered it without
+  being told to, which is the point of reading the registry instead of a list.
+
+  No version bump -- this is documentation and tests only.
+
+## [3.0.0] — 2026-09-30
+
+### Changed
+
+- **BREAKING: `confirmed` is now `executed`, and three settled measurements
+  stopped reporting as `needs-review`.** Every verdict is now named for what the
+  *target* did, never for how sure RCEKit is.
+
+  Two problems, one vocabulary. `confirmed` was the only verdict that described
+  RCEKit's own confidence while every sibling described the target's behaviour,
+  and next to it a `lookup-sink` -- proven on a callback carrying a token only
+  that probe held -- read as something unconfirmed. And `needs-review` had four
+  users, of which three had measured their answer:
+
+  | Verdict | Was | What it says |
+  |---|---|---|
+  | `executed` | `confirmed` | the target executed the input |
+  | `timing-sink` | `needs-review` | the target honoured a delay RCEKit injected |
+  | `file-write` | `needs-review` | the target stored a file at a path RCEKit chose and serves it back uninterpreted |
+  | `evaluation-sink` | `needs-review` | an evaluator consumed the input and partitioned on it; which evaluator is not shown |
+  | `needs-review` | `needs-review` | unchanged, and now with one user: `deser`'s parser fingerprint, the one signal that really is a candidate |
+
+  `write`'s middle verdict is the clearest case: its evidence line opened with
+  the words "ARBITRARY FILE WRITE confirmed" under a status that said
+  `needs-review`. The tool's own prose contradicted the tool's own status,
+  because the status was naming a confidence the measurement did not have a
+  reason to lack -- the file was on the target and RCEKit had read it back.
+
+  Nothing was promoted. `executed` still rests on exactly the three channels it
+  did before, and none of the five weaker tiers has a path into it. `time` in
+  particular stays out: a delay honoured is not necessarily a *shell* honouring
+  it, since a sandbox that implements `sleep` answers the regression the same
+  way. What changed is the direction of the reporting error -- three proven
+  findings were being under-reported.
+
+- **`overall_detection_verdict` ranks `needs-review` last instead of second.**
+  Its position was argued from "a suspected RCE outranks a proven non-RCE in
+  triage", which held while `time`, `boolean` and `write`'s uninterpreted file
+  all reported there. The one verdict left under it is a suspected
+  *deserialization*, so ranking it above `deserialization-sink` put the guess
+  above the proof of the same thing. The full order is now `executed`,
+  `timing-sink`, `file-write`, `evaluation-sink`, `deserialization-sink`,
+  `lookup-sink`, `needs-review`.
+
+  That order is stated once, in `NON_EXECUTION_SECTIONS`, and both the
+  collapsed verdict and the text report read it from there. They used to state
+  it separately -- one as a tuple, the other as the order of print statements
+  in `main()` -- and they disagreed: `deserialization-sink` and `lookup-sink`
+  were hand-written blocks above the table holding the rest, so a run
+  reporting both `timing-sink` and `deserialization-sink` collapsed to the
+  first and led the report with the second. A statement order is not data and
+  cannot be checked against anything.
+
+  Folding those two in unifies their heading with the others: `NOT proof of
+  RCE` becomes `NOT proof of execution`, which is what the other four already
+  said. Their rows keep the compact form they had -- truncated payload, no
+  `environment` -- because they carry base64 object streams and JNDI URIs.
+
+- The classic `--verify-url` path renamed its `confirmed` verdict too, and its
+  section headings with it. It has no sinks and no `needs-review`, so nothing
+  there was mis-tiered; a tool printing `[verify] CONFIRMED` beside `[detect]
+  EXECUTED` would just be two vocabularies. That path has no machine-readable
+  output, so this is text only.
+
+### Benchmark
+
+**`runner.py --all`: 9/9, in 55m50s**, against the same vulhub builds. Every
+case's expected verdict changed in this release, and a rename applied to a
+stored result proves nothing about what the tool now reports, so all 9 ran:
+
+| Case | Method | Vulnerable half | Negative control |
+|---|---|---|---|
+| `fastjson-1.2.83` | `deser` | `deserialization-sink` | `negative` |
+| `hugegraph-gremlin-eval` | `eval` | `executed` | `negative` |
+| `hugegraph-gremlin-shell` | `reflected` | `executed` | `negative` |
+| `log4shell-cve-2021-44228` | `lookup` | `lookup-sink` | `negative` |
+| `opentsdb-cve-2023-25826` | `oob` | `executed` | `timing-sink` |
+| `struts2-s2-001` | `eval` | `executed` | `negative` |
+| `tomcat-cve-2017-12615` | `write` | `executed` | `negative` |
+| `webmin-cve-2019-15107-file` | `file` | `executed` | `negative` |
+| `webmin-cve-2019-15107` | `reflected` | `executed` | `timing-sink` |
+
+The two `timing-sink` controls carry the weight. On a Webmin 1.910 and an
+OpenTSDB 2.4.1 that are genuinely vulnerable, `time` reaches its own proven
+tier and still does not reach `executed` -- the retier holding against real
+software and not against a fixture.
+
+Two tiers this release introduced are **not** in that table: `file-write` and
+`evaluation-sink`. The Tomcat `write` case reaches `executed`, and `boolean`
+has no case at all. Both are covered by unit tests only.
+
+### Migration
+
+`--detect-json` changes values, not keys. `verdict`, the keys of `counts`, each
+probe's `verdict`, and `observe_status` take the new names per the table above.
+A consumer keyed on `"confirmed"` must read `"executed"`; one keyed on
+`"needs-review"` should decide which of the four it actually meant. No CLI flag,
+no JSON key, no template-schema field and no exit code changed.
+
+No backward-compatible alias is shipped. Two accepted spellings for one field is
+a second source of truth, and the point of this release is that there is one.
+
+### Fixed
+
+- **A method's question was derived from its tier, which does not determine
+  it.** `detection_question` decided whether a method asks *did the target
+  execute my input* by looking up its tier in a set of two names. A tier says
+  how strong an answer a method can reach; `time` asks the execution question
+  and cannot reach the strongest answer to it, while `lookup` reaches a
+  definitive answer to a different question.
+
+  Left alone, renaming `time`'s tier in this release would have moved it out of
+  the execution group and out of the dedup that skips it on a candidate
+  `reflected` has already proven -- so the most expensive method in the tool
+  would have run to put a second name on a finding. `question` is now declared
+  on the class, `EXECUTION_TIERS` is gone, and the partition is unchanged.
+
+  The test covering this recomputed the derivation rule it was checking, so it
+  agreed with the rule whatever the rule became. Three tests replace it: the
+  partition written out, the one case where question and tier disagree, and the
+  `time`-leaves-the-group regression by name.
+
+## [2.45.7] — 2026-09-26
+
+### Fixed
+
+- **The `write` product overflowed a 32-bit int, so the method reported
+  `negative` on more than half of the targets it was built for.** RCEKit
+  computes `a*b` in Python, where integers do not overflow. The interpreter on
+  the other side usually is not Python: JSP and ASPX are 3 of the 5 languages
+  this method writes, and both evaluate the expression as a signed 32-bit int.
+  A product past 2**31-1 therefore came back wrapped, disagreed with the value
+  RCEKit had computed, and the run reported `negative` against a target that is
+  fully exploitable -- the one failure this tool exists not to produce.
+
+  Measured against Tomcat 8.5.19 (CVE-2017-12615), writing the probe by PUT and
+  fetching it back. `97233*38786` was served as `-523688158` where RCEKit
+  expected `3771279138`; `4721*8093` came back as `38207053`, exactly. The
+  operand range drew a pair whose product overflows in **56%** of runs, and the
+  operands are drawn once per run, so there was no second probe to recover: more
+  than half of all runs against a JVM or .NET target lost.
+
+  The operands are now 4 digits apiece, which caps the product at 99,980,001 --
+  the range `eval` has always used against these same targets. Nothing about the
+  oracle changes: the product is still bracketed by two random tags and still
+  differenced against a payload-free control, which is what makes it unforgeable
+  rather than its magnitude.
+
+  The same run confirms now. `RKYBRDH<%=3253*8793%>RKICSMR` came back as
+  `RKYBRDH28603629RKICSMR` from the fetched file, and `write` reached
+  `confirmed` against real software for the first time.
+
+  Two tests, both written first and watched to fail: the drawn operands multiply
+  to something a signed 32-bit int can hold, across 400 seeds; and a store whose
+  evaluator wraps to 32 bits the way a JVM does reaches `confirmed` rather than
+  `negative`. A third test had pinned the operand width with a hard-coded
+  `\d{5}`; it reads the declared range now and asserts the operands fall inside
+  it, which the digit count had only implied.
+
+## [2.45.6] — 2026-09-26
+
+### Fixed
+
+- **The `write` probe carried whitespace, and some sinks tokenise before they
+  write.** The file content this method delivers is a language one-liner, and
+  its templates spelled the expression with spaces around it -- `<?= a*b ?>`,
+  `<%= a*b %>`. The content goes through the operator's own injection point, and
+  a sink that splits its input on whitespace never writes a valid file.
+
+  Measured against RRDtool 1.7.2, reached through Cacti 1.2.28's
+  `right_axis_label` (CVE-2025-24367), which builds the file from a
+  `LINE1:out:<content>` argument. `RKAAA<?=7*6?>RKBBB` wrote a 47-byte file with
+  the content intact. `RKAAA<?= 7*6 ?>RKBBB` wrote no file at all and answered
+  `ERROR: '7*6' is not a valid function name in 7*6`. The whitespace was the
+  whole difference.
+
+  Whitespace inside `<?= ?>` and `<%= %>` is optional, so the templates drop it:
+  nothing is lost and every sink of that shape is gained. Verified end to end
+  against the real sink rather than a fixture -- RCEKit generated
+  `RKYOOQS<?=69294*83370?>RKGFZQP`, RRDtool wrote it whole, and PHP answered
+  `RKYOOQS5777040780RKGFZQP`, which is the value RCEKit had computed for that
+  probe.
+
+  `jspx` keeps its spaces: it is an XML document whose root element carries
+  namespace attributes, and a sink that tokenises on whitespace was never going
+  to carry one. Two tests pin the rest -- the generated probe and the templates
+  themselves -- both written first and watched to fail.
+
+## [2.45.5] — 2026-09-25
+
+### Fixed
+
+- **The `deser` shape differential answered from evidence it had not
+  validated.** Three ways, each found by pointing the method at a live
+  fastjson 1.2.83 endpoint under vulhub and asking why it said what it said.
+
+  **Drift landed on whichever form went last.** The differential is read across
+  requests, so anything changing with the request *index* rather than with the
+  payload -- a rate limiter backing off, a filling log, a warming cache -- fell
+  on the form sent last, and noise was sent last every time. There are two
+  noise probes now, first and last, bracketing the structured pair. They are
+  one noise body sent twice, byte for byte, because the bracket asks whether
+  the endpoint answered *the same probe* alike and anything the target can see
+  a difference in it may answer a difference to: a tail beginning `n` made
+  fastjson say "error parse new" where every other tail said "syntax error".
+  A carrier whose two ends disagree is held at `inconclusive`. The engine
+  de-duplicates by payload on the per-probe path but not on the aggregate path
+  where `deser` lives, which is what makes one body twice possible; a test pins
+  that, because the bracket becomes a single request if it ever changes and a
+  drift check comparing an answer with itself passes forever.
+
+  **An endpoint echoing its input looked like a parser.** A response carrying a
+  fragment of the payload that begins at the format's magic, and not carrying
+  it in the payload-free control, is an echo and the shape channel cannot be
+  read through it -- `inconclusive`, not `negative`, because the channel was
+  unreadable rather than empty. The sentinels are anchored on the magic's
+  position in the payload rather than its start, since a probe is wrapped for
+  its injection context first and some wrappers are longer than the magic: an
+  `xml_cdata` pickle probe begins `<![CDATA[gASV`. They are matched raw, case
+  folded (because `_signature` folds case and the two halves of one oracle must
+  agree), and encoded -- base64 aligned on bytes rather than characters and
+  read from the first three-byte boundary at or past the magic, hex from the
+  magic itself.
+
+  **A format declaring no magic got a verdict anyway.** Without a magic there
+  is no sentinel, so an echoing endpoint cannot be told from a parsing one --
+  and the differential answers regardless, because the noise form is random
+  where a structured pair is not. Every shipped format declares one; a
+  `--template-file` need not, and the answer to a control that cannot be
+  validated is to say so.
+
+- **The page an operator plans traffic from still described three payloads.**
+  The bracket makes the shape oracle send four per ecosystem -- 20 requests per
+  carrier rather than 15, with the five the corpus ships -- and
+  `docs/reference.md` and the method's own docstring both still said three.
+  README advertises an execution plan that prints the exact probe count before
+  anything fires, so a page that undercounts it breaks that promise in the one
+  place someone reads before a monitored engagement.
+
+  Both now state four, and the arithmetic, and what the repeated probe is for.
+  A test holds the page to `SHAPE_FORMS` and to the ecosystem count, in both
+  directions: adding a form without updating the page fails it, and reverting
+  the page without touching the code fails it too. Nothing compared the two
+  before, which is how the count drifted the moment the form was added.
+
+- **The `negative` evidence line described a comparison it had not made.** "The
+  endpoint answers all three forms alike" was asserted rather than observed.
+  It now names what actually collapsed: every form alike, a truncated stream
+  answered as a complete one, or the noise form answered as a structured one.
+
+### Withdrawn before release
+
+- **A second route through the differential, reading noise as the control.** It
+  was written for the shape fastjson has -- a format resolving its type name
+  before the parse finishes answers a truncated stream exactly as it answers a
+  complete one, which collapses the original comparison on the very endpoints
+  it exists to find -- and it did reach `needs-review` there.
+
+  It is not in this release. "Noise differs from the structured pair" has too
+  many explanations that are not a parser. Drift and echo are closed above. The
+  third is not: a filter that refuses the random payload while letting the
+  structured ones through produces the same differential, no repetition exposes
+  it because the refusal is stable, and this verdict cannot see it -- refusal
+  is judged against the probe the verdict is anchored to, and that is the
+  well-formed one.
+
+  A format of that shape is reported `negative` by the shape differential and
+  reaches `deserialization-sink` through the DNS gadget instead, which is this
+  method's real oracle and rests on a callback rather than on the shape of a
+  response. Measured on the same target: 14 callbacks, each carrying a token it
+  could only have learned by resolving the name it was handed.
+
+## [2.45.4] — 2026-09-25
+
+### Fixed
+
+- **A compressed response body is a body the oracles cannot read, so a target
+  that executed the probe was reported `negative`.** Every in-band oracle --
+  `reflected`, `eval`, `file`, `write`, `boolean` -- confirms by finding a
+  value the target computed in the response. Nothing in the delivery layer
+  undid `Content-Encoding`, so a gzipped body arrived as bytes that *contain*
+  the computed value and do not spell it, and the search for a decimal number
+  came back empty. The verdict was `negative`: not "could not read this", but
+  "the probes reached the target and it is not vulnerable".
+
+  Measured against Apache HugeGraph 1.2.0 under vulhub. The probe
+  `echo RKPDSWY$((258675+956113))RKIEMTB$(echo RKLUQKL)RKPDSWY`, delivered
+  through the Gremlin API, came back HTTP 200 carrying
+  `RKPDSWY1214788RKIEMTBRKLUQKLRKPDSWY` -- the sum and the substitution
+  collapse, both of `ReflectedMath`'s proofs. Present after gunzip, absent in
+  the raw bytes, and the run reported `negative` across 620 probes.
+
+  What made it survive this long is that the same target confirmed on the same
+  run's *other* probes: HugeGraph gzips its 200s and leaves its 4xx errors
+  plain, so `eval` confirmed through a 400 that carried a Groovy exception
+  message while the 200 that carried real command execution was invisible. A
+  run that finds something is a run nobody rereads.
+
+  Both HTTP entry points decode now -- the verification/detection path and the
+  `--verify-chain` path -- through one helper, on the success body and the
+  `HTTPError` body alike, since an evaluator surfacing its value in a 500 is
+  the case that path exists for. The header is honoured whether or not RCEKit
+  asked for it, because HugeGraph compresses a response to a request that sent
+  no `Accept-Encoding` at all.
+
+  `gzip` and `deflate` only, both from the standard library, and `deflate`
+  tries the zlib wrapper before the raw stream because servers disagree about
+  which one that name means. `br` and `zstd` need a third-party module and this
+  tool has none to add; a body in either is returned as its raw bytes. Nothing
+  here raises: a body that claims an encoding it is not in comes back
+  undecompressed rather than ending a run that has already spent hundreds of
+  probes.
+
+  This does not widen what `confirmed` means. The oracle is unchanged -- a
+  computed value present in the response and absent from a payload-free
+  control -- and nothing is promoted into it. What changes is that the response
+  the oracle reads is the response the target sent.
+
+- **Decompression is bounded, because compression is where the target chooses
+  how much memory RCEKit allocates.** Reading a body at all is new here, and
+  one-shot decompression handed a blank cheque to the one party in the exchange
+  that is untrusted by definition -- the thing being tested. Measured: 130,478
+  bytes of gzip expand to 134,217,728, and 203,860 bytes hold 200 MB of zeros,
+  a ratio of 1:1028. This path runs once per probe.
+
+  The ceiling is 32 MiB per encoding step. Whatever was produced before it is
+  kept rather than discarded: the value the oracle wants is short, so searching
+  32 MB is a far better answer than searching nothing, and the only risk the
+  ceiling carries is a false negative -- never a false `confirmed`. A truncated
+  read is logged rather than passed off as a clean one, since a `negative`
+  decided on a partial body is not the same claim as a `negative` decided on
+  the whole of it.
+
+  Multi-member streams are still followed. `gzip.decompress` followed them, and
+  a ceiling that quietly became a truncation for a server that concatenates
+  members would be the same false negative in a new place.
+
+## [2.45.3] — 2026-09-24
+
+### Fixed
+
+- **A `deser`-only run started no listener, so its gadgets went out with
+  nothing to receive them.** `--methods deser --oob-host <delegated-name>`,
+  with neither `oob` nor `lookup` selected, built its two DNS-gadget probes and
+  sent them to a target while no listener was running -- with the host
+  configured perfectly. `deser` reached `deserialization-sink` only when it
+  happened to be selected beside a method that required a host; run on its own,
+  the way the docs present it, its strongest oracle was structurally dead and
+  it capped at `needs-review` without saying so.
+
+  The gate read `needs_oob_host`, which was answering two questions at once:
+  "should the run stop when no host is named" and "does this run need a
+  listener". Every other method answers both the same way. `deser` is where
+  they come apart -- its shape oracle proves something with no listener at all,
+  so a missing host must not stop the run, while its DNS gadget does call back
+  and needs one.
+
+  So the second question gets its own attribute, `uses_oob_host`, and the
+  listener starts for any selected method that calls back once a host is named.
+  It is declared rather than derived: `needs_oob_host or oob_needs_dns_label`
+  gives the right answer for every method shipping today and would miss one
+  that called back over HTTP without requiring a host -- which is the shape
+  `oob` itself would have if its host were optional.
+
+  Two notices could now reach a run with no HTTP shape in it, and said "only
+  the HTTP shapes are live" as though `oob`'s channel were the only one. They
+  say "only the shapes that need no DNS are live", which is true of `oob`'s
+  HTTP probes and of `deser`'s shape oracle alike. The message for an
+  unbindable HTTP port no longer names `lookup` as the reason a run continues.
+
+  `uses_oob_host` is the default answer rather than the answer. Whether a
+  method calls back can depend on its own flags: `--deser-formats` selects the
+  ecosystems, and three of the five -- `php`, `dotnet`, `python_pickle` --
+  ship no DNS gadget at all. So both callers ask `builds_callback_probes()` on
+  a configured instance instead of reading the class. `--deser-formats php`
+  with an address no longer draws a notice blaming the address, because there
+  was no callback probe to lose and a delegated domain could not have helped;
+  and it no longer starts a listener that nothing would reach.
+
+  Whether the host's *form* is one a method can use is a second question,
+  `can_use_oob_host()`, and only the gate asks it. A listener bound for
+  `--methods deser --oob-host 10.0.0.9` receives nothing -- `build_probes`
+  drops an address and sends no gadget -- and the run announced "the target
+  will open outbound connections" directly beneath the notice saying those
+  probes are not sent. It stays out of `builds_callback_probes()`, because the
+  stranded notice is *about* the address and an address test there would
+  silence it for the one case it exists to report.
+
+  `lookup` with an address stops getting a listener too. It built no probes
+  with one either way, so nothing is lost, and the stranded notice says more
+  about that run than the port-53 warnings it no longer reaches.
+
+  A flaky test met on the way, fixed here because it would have bitten CI
+  sooner or later: the fake parsing endpoint behind the `deser` shape
+  differential sniffed one byte (`raw[4:5] not in (b"t", b"s")`) to decide
+  whether a stream was well-formed. The `noise` form is the format's magic
+  followed by random characters, so it passed that check about once in 500
+  runs -- measured at 398 in 200,000 -- all three forms answered alike, and
+  the oracle's correct `negative` failed a test that wanted `needs-review`.
+  The endpoint now reads TC_STRING the way a parser does, a two-byte length
+  followed by exactly that many bytes: 0 collisions in 500,000 seeds. Seeding
+  the test's RNG would have hidden it instead, and would have gone on hiding a
+  real regression in how the noise form is built.
+
+  This is the second half of #94, found while fixing the first and reported as
+  #98 rather than folded into #97: same defect -- probes sent that no callback
+  could follow -- reached by a different route. The format-dependent part came
+  from an automated review of #97 and the host-form part from one of this
+  change, both verified before being accepted.
+
+## [2.45.2] — 2026-09-24
+
+### Fixed
+
+- **`deser` fired DNS gadgets at a host that could not resolve, and the warning
+  built for exactly that stayed quiet.** `--oob-host` takes a delegated name or
+  an address. `oob` uses either: given an address it drops its DNS shapes and
+  carries the token in a URL path instead. `lookup` refuses an address
+  outright, because a lookup expression has nowhere else to put one.
+  `DeserSink` had neither behaviour -- it built `<token>.10.0.0.9` and sent it,
+  a real request the target answered from which no callback could follow by
+  construction, and then capped at `needs-review` without saying why.
+
+  `oob_channel_warnings` is the function that exists to say a DNS channel is
+  structurally dead. It returned early on an address literal, reasoning that
+  such a run "never builds DNS probes and has nothing to warn about" -- true of
+  `oob`, and written when `oob` was the only method taking the flag. So the one
+  mechanism that could have reported this was the one place it was ruled out.
+
+  Both halves are fixed together, because either alone still ships the problem.
+  `deser` treats an address as no host and sends its shape oracle alone, which
+  is what it can prove without a listener anyway. The notice moves into
+  `oob_address_strands`, which names the methods an address stranded and the
+  tier that puts out of reach, read off each class.
+
+  It is a separate function rather than another branch of
+  `oob_channel_warnings` because the two answer different questions at
+  different times. "Can the DNS shapes reach this listener" is only worth
+  asking once a listener exists, and that block runs only for a method that
+  *requires* a callback host. `deser` does not require one, so it never
+  reaches that block -- a notice left there could not have fired for the
+  method that prompted it. The new check runs on every verification run that
+  names a host.
+
+  Which methods those are is a class attribute now, `oob_needs_dns_label`,
+  declared beside `needs_oob_host`. Deciding it at the one place that asked is
+  how the two drifted apart to begin with.
+
+  The test that pinned the old silence asserted "no DNS shape is ever built" --
+  a claim about `oob` wearing the shape of a claim about the flag. Its scope
+  narrows to `oob`, and the stranded case gets tests of its own.
+
+  Found while verifying an automated review finding on #93; reported as #94 and
+  deliberately kept out of that documentation-only PR.
+
+## [2.45.1] — 2026-09-24
+
+### Fixed
+
+- **`--max-payloads` is a bound on requests, and it was being checked after
+  they had been sent.** The cap was measured against the number of result
+  *rows*. For a method that answers from each probe those are the same number,
+  so nothing showed. An aggregate method reports one row however many probes it
+  cost, so the whole series went out and the cap noticed afterwards:
+
+  | `--max-payloads 1` | the cost line said | requests sent |
+  |---|---|---|
+  | `reflected`, `eval` | 1 | 1 |
+  | `time` | 1 | **12** |
+  | `deser` | 1 | **15** |
+  | `boolean` | 1 | **27** |
+
+  A clean split: every aggregate method overran and no per-probe method did.
+  That line is the only thing an operator bounding a monitored engagement has
+  to go on before the traffic starts, and it was wrong in the direction that
+  matters — under, not over.
+
+  Rows were the wrong meter in a second way, which the first fix exposed. A
+  series the budget declines costs requests and produces no row at all, so the
+  next carrier recomputed the same allowance and fired again — `time` at
+  `--max-payloads 5` sent 4 requests per carrier with the cap never moving. And
+  in a mixed run, a per-probe method never saw what an aggregate one had
+  already spent: `boolean,eval` at 30 sent 56. The budget is counted in
+  requests for the whole call now. Swept across every method this build can run
+  without a callback host, alone and in combination, at seven caps: 63
+  combinations, worst overrun 0.
+
+- **A measurement that cannot be finished is no longer started.** What a budget
+  may do to a series depends on where the method's answer lives, and the class
+  already knows: `DetectionMethod.decides_per_probe()` reads it from whether
+  the class overrides `confirm_each`, rather than from a list of names that
+  would go stale like every other one in this repository.
+
+  Where each probe answers — `deser`, `lookup`, `oob` — running out of budget
+  stops the series and the probes already sent keep their verdicts. Where only
+  the series answers — `time`, `boolean` — a part of one is not a weaker answer
+  but a wrong one: `time` reports `negative` from a screen with no regression
+  behind it and from a regression short of four samples, and a `boolean` series
+  whose anchors never went out reads the same way. Both are false cleans, and
+  the budget would have been manufacturing them. So a wave that does not fit
+  abandons the measurement rather than cutting it short.
+
+- **An abandoned measurement leaves an `inconclusive` row, not a silence.**
+  Dropping it quietly let the *other* carriers describe the run, and the other
+  carriers are the ones with nothing to find: measured at `--max-payloads 12`
+  against a sink that honours an injected sleep, the unix carrier's regression
+  was abandoned for budget and a windows carrier's honest "no separator
+  delayed" was the only row left — so the run reported `negative` for a target
+  that was vulnerable. `inconclusive` is what an abandoned measurement is in
+  the word the tool already uses, and it outranks `negative` in the run
+  verdict. Swept across 31 caps against a sink that really does delay: no cap
+  reports it clean, and none overruns.
+
+- **What the budget declined is reported by name**, in a sixth run-wide tally
+  beside the profile drops, the safety holds, the reach notes, the settled
+  carriers and the refusals. A ladder that shrinks quietly is indistinguishable
+  from a target with nothing to find:
+
+  ```
+  [detect] --max-payloads held back 1 measurement(s) that could not have reached a
+           verdict within the budget:
+  [detect]   1 x boolean/raw needs 7 requests to reach a verdict and --max-payloads left 3
+  [!] No probes were built, so NOTHING WAS TESTED - this is not a negative result.
+  ```
+
+- **The cost line and the run make the same decision.** `estimate_detection_probes`
+  applies the same all-or-nothing rule, so under a cap the figure printed before
+  the traffic is the traffic. Uncapped it stays a floor by documented design — a
+  wave a method picks after seeing its own timings cannot be predicted from
+  there.
+
+- **`sent N probes` is a claim about what the target received**, and it was
+  counting result rows. Those are the same number for a method that answers
+  from each probe, which is why it read true for so long -- but an aggregate
+  method reports one row for a whole series, so a `time` run that put 20
+  requests on a target announced 5, and a measurement this release declines
+  would have announced one probe for traffic that never left. Both the run
+  summary and the per-point line report deliveries now, with the result count
+  beside them.
+
+  One figure in the 2.45.0 entry below came from that line and was mislabelled
+  with it. Re-measured against the same sink: `--methods reflected,eval,time`
+  put **2883** requests on the target, not 2426, which was the number of
+  results. Every one of them was still `negative`, so the claim stands; the
+  number did not.
+
+- **An evasion retry is a request, and the meter could not see it.** `_escalate`
+  fired without incrementing `delivered_probes`, so retries were traffic
+  outside the bound this release makes real: measured at `--max-payloads 5
+  --evade high` against a filter that refuses whitespace, the target received
+  12 requests and the run recorded 5. Counting them was half the fix. The other
+  half is that the budget is checked *per rung* rather than once before the
+  ladder -- one refused probe is retried at `low` and again at `high`, so it
+  can cost three requests, and a check at the call site alone still overshot by
+  one at `--max-payloads 10 --evade high`. Swept over caps 1 to 25 at all three
+  evasion settings: 75 combinations, no overrun, and the counter equal to the
+  requests the target actually received in every one.
+
+### Security
+
+No authorization, deny-by-default or matrix-evaluation boundary moves, and
+nothing widens `confirmed`. The effect is one-way: fewer requests than before,
+and a run that could not afford to test something now says so instead of
+reporting it clean.
+
+## [2.45.0] — 2026-09-23
+
+### Added
+
+- **`--methods boolean` — a sink that evaluates a predicate and renders nothing
+  of it.** MongoDB `$where` is the shape, and this repository has carried the
+  note for two releases: a JS sandbox with no shell, no egress and no value in
+  the response, only a document set that a predicate narrows. Every shipped
+  oracle is structurally blind to it. Measured against exactly that sink:
+
+  ```
+  [detect] methods: reflected, eval, time
+  [detect] sent 2426 probes: negative=2426
+  ```
+
+  2426 requests and a clean negative on a target that evaluates whatever it is
+  handed. The new method reads the one channel left — whether the *shape* of the
+  response changed between a true predicate and a false one.
+
+  **It is `needs-review` and there is no path from here to `confirmed`.** Not
+  because the signal is weak, but because of what it cannot distinguish: against
+  a sandboxed `eval` sink and against a plain SQLite comparison it produced an
+  identical clean differential in 40 runs each, and a query engine comparing two
+  numbers is not remote code execution. Extracting a locally computed product
+  bit by bit through the channel was tried and does not fix it — it recovers the
+  product through both sinks alike, for about 80 requests and a string function
+  a sandbox may well deny.
+
+  **The naive form of this oracle is unusable**, which is why none of it is.
+  `1==1` against `1==2`, with a changed response read as a finding, called a
+  target that only *reflected* its input vulnerable in 40 runs out of 40, and
+  one whose response merely wobbled in 32 of 40. Four guards, each a measured
+  false-finding rate rather than a precaution:
+
+  - **Compare structure, not the body and not its length.** A reflected payload
+    lands in the text between two tags, and the text between two tags is what
+    the signature throws away. A length-based signature claimed a differential
+    in 13 of 25 runs against a reflect-only target; comparing raw bodies was
+    unusable outright, reading `unstable` in 25 of 25 runs against a target that
+    *was* vulnerable, because one CSRF token makes every response unique.
+  - **Several independently randomised pairs, not one.** Against a target whose
+    response varies on its own, one pair claimed a differential in 46 of 200
+    runs; two claimed none in 200. `--probe-depth quick` trades four pairs for
+    two and never for one.
+  - **Randomised firing order.** A target that never reads the payload but
+    degrades part-way through a run splits an ordered true-then-false series
+    perfectly: at the worst point of a swept degradation, 100 false findings out
+    of 100.
+  - **An anchor before and after the series, each a different true predicate.**
+    Shuffling alone still left 2 in 100, which is just the chance a shuffle
+    lands separable. Re-measuring the channel afterwards caught it 100 times in
+    100, because a target that moved during the series cannot answer the
+    closing anchor the way it answered the opening one. Sending *one* anchor
+    payload three times does not merely weaken that: a cache keyed on the query
+    string answers the repeats from its store, so the closing anchor agrees
+    with the opening one whatever the target did in between. Measured against
+    an input-blind target that degrades mid-series, identical anchors caught it
+    in 36 of 39 runs live and in 0 of 39 behind a cache. Every probe payload is
+    unique, so nothing else in the series is replayable.
+
+  With every guard on, a genuinely evaluating target still read as a
+  differential in 100 runs of 100.
+
+- **A channel that cannot carry one bit is `inconclusive`, never `negative`.**
+  If the same probe draws two different shapes, or the shape moves while the
+  series is being fired, the run says so. `negative` asserts the probes reached
+  the target and found nothing; here they reached it and no answer could be read
+  out of them, which is the same false clean `blocked` and `nothing-tested`
+  exist to prevent, one oracle further in. No new verdict: there are still nine.
+
+- **`response_shape()`** — a response reduced to its structure, with everything
+  it said removed. Three readings, because a response is one of three things and
+  the wrong reading is not a near miss: a JSON document keeps its keys, nesting
+  and list lengths and drops every scalar; markup keeps its tag skeleton;
+  anything else keeps one marker per word per line. Measured on a JSON sink, the
+  markup reading was unusable — `unstable` in 25 of 25 runs — and the shape tree
+  read the differential in 25 of 25.
+
+### Changed
+
+- **`CODE_POSITION_CONTEXTS`** names the contexts that carry the injected value
+  as code rather than as a value, and `boolean` is offered every other one. The
+  first cut of this asked whether the context had a break-out prefix at all, and
+  got both halves wrong: it refused `attribute`, `attribute_unquoted`,
+  `xml_cdata` and `yaml`, whose delimiters open and close *around* the value and
+  leave a predicate exactly where a predicate belongs, and it offered
+  `unix_shell`, `windows_cmd` and `powershell`, which have no delimiters at all
+  and run the value as a command. The suite enumerates both sides, so a context
+  added to the corpus fails until somebody decides which one it is on.
+
+- **`costly` now asks whether one probe buys an *answer*, not whether it costs
+  more than one response.** The two were the same question while every method's
+  probe was also its unit of information. `boolean` is the first where they come
+  apart: each of its probes is one ordinary request and none of them means
+  anything alone, because the answer is the partition across the whole series.
+  Read the old way it would have landed in the wave the enumeration driver runs
+  *first* — the one that exists to be answered cheaply — ahead of `reflected`
+  and `eval` and spending the same per-question budget, at 27 requests before it
+  could say a word. It sits with `time` instead, and the suite now holds every
+  aggregate method to that.
+
+### Security
+
+- **The `OR` connectives ship behind `--verify-active-risk stateful`.** A
+  predicate probe breaks out of a condition the application already wrote, and
+  the connective is this method's command separator. `AND` differentiates only
+  where the application's own predicate is true and `OR` only where it is false,
+  so they are complements and dropping `OR` is a blind spot rather than a
+  saving. But a true predicate `OR`-ed into a `DELETE … WHERE` took a table from
+  3 rows to 0, where the same predicate `AND`-ed into it left all 3 — so it goes
+  at the top rung, held back by default, with the run naming every shape it held
+  and the flag that sends it. `AND` and the bare form stay `safe` and change
+  nothing.
+
+- **Nothing widens `confirmed`.** The new method's ceiling is one tier below it
+  and the suite holds that as behaviour, not as an attribute: `confirm_series`
+  is driven across 200 series including the perfect one, and none of them
+  reaches `confirmed`.
+
+## [2.44.0] — 2026-09-23
+
+### Fixed
+
+- **`--evade low` substituted inside quoted programs and broke them.** Every
+  space became `${IFS}`, including the ones inside `awk 'BEGIN{print "RK" a+b
+  "RK"}'` — and inside single quotes `${IFS}` is literal text, not an
+  expansion, so awk was handed `BEGIN{print${IFS}"RK"...` and answered with a
+  syntax error.
+
+  Measured shape by shape against an unfiltered target: **8 probe shapes that
+  the canonical form executes broke at the rung, and none improved.** The
+  substitution stops at a quote now. Double quotes are left alone from the
+  other side — `${IFS}` *does* expand inside them, so substituting there would
+  change the string the target computes rather than the spacing around it.
+
+### Changed
+
+- **A rung is a retry for a refused probe, not a posture for the run.** It was
+  applied to every probe regardless of whether anything was being filtered,
+  which is a pure loss on a target with no filter. Against a filter that blocks
+  whitespace it turned 1 confirmation into 5; against an unfiltered target it
+  now costs **zero** extra requests, because nothing was refused.
+
+  `--evade` is therefore a **ceiling**. Every probe goes out canonical, and
+  only a refused one is retried, up to that ceiling — at most one request per
+  rung. The ladder a run builds no longer depends on the rung at all: all three
+  settings build the same 42 probes, of which the same 32 execute. Before, the
+  rung built a smaller ladder and 13 fewer of its probes ran.
+
+  The run reports how many retries it made and which rung got through, whatever
+  it concluded — including a run that confirmed, where the retry is the reason
+  it did.
+
+### Added
+
+- **A second evasion rung, `high`.** The measurement turned up two classes of
+  filter and the shipped rung only addressed one. `low` removes whitespace;
+  `high` also splits the command word with an expansion that vanishes
+  (`ec$@ho`), for a filter matching command names. It is applied before the
+  whitespace substitution, because afterwards the split lands inside `${IFS}`
+  and makes `${I$@FS}` — neither an expansion nor a command.
+
+  Four boundaries the retry has to respect, each of which was a way to turn a
+  vulnerable target into a `negative` — worse than the `blocked` the rung sits
+  beside, because `blocked` at least says the run learned nothing:
+
+  - **Unix shell probes only.** `${IFS}` and `$@` are POSIX. A cmd.exe probe
+    rewritten with them loses its spaces, so a whitespace filter answers 200
+    and the retry counts as a win while cmd.exe cannot run it.
+  - **A break-out context opens by *closing* a quote.** Reading that leading
+    quote as an opener left `'; echo …` untouched, so the rung did nothing on
+    exactly the contexts a filter is most likely to sit in front of.
+  - **A followup is read again after a retry that lands.** `file` writes its
+    token on the request that arrives, so a body read before the retry is a
+    read of a file that did not exist yet.
+  - **A redirect is never retried.** The build-time transform took an explicit
+    `evade=False` for these, and that parameter stopped doing anything when the
+    rung became a retry — a guard lost in the move, restored where the retry
+    now happens.
+
+  The aggregate methods escalate too. `time`, `oob`, `lookup` and `deser` take a
+  different branch, and leaving it out meant the documented ceiling did nothing
+  for four of the eight methods — the same branch, and the same omission, as the
+  refusal check one change earlier.
+
+## [2.43.0] — 2026-09-23
+
+### Added
+
+- **A ninth verdict: `blocked`.** A run whose payloads a filter refuses has
+  learned nothing about the sink, and it used to say otherwise. Measured
+  against a real command injection behind a filter that 403s a space or a
+  separator:
+
+  ```
+  before:  sent 10 probes: negative=10
+  after:   sent 10 probes: blocked=10
+  ```
+
+  `negative` asserts that the probes **reached** the target — the project's own
+  tier rule says so. They reached a filter. That is the same false clean
+  `nothing-tested` exists to prevent, one level further in.
+
+  **The signal is differential**, like everything else this tool decides: the
+  payload-free control got through and the probe did not, so what was refused
+  is the payload. An endpoint answering 403 to everything — an auth wall, a
+  path that does not exist for this session — refuses the control too and is
+  not mistaken for a filter. Verified against exactly that case.
+
+  4xx only. A 5xx is as likely to be the payload *breaking* the application,
+  which means it reached something, and reading that as blocked would hide the
+  one response saying the sink is live. No vendor list and no block-page
+  fingerprints: a status the control did not get is the whole signal.
+
+  A run where **some** probes got through stays a real `negative` — the sink
+  saw those — and the refusals are reported either way.
+
+  **A refusal never unmakes evidence.** Only a `negative` is replaced, because
+  it is the only verdict a refusal contradicts. An application can execute the
+  payload and then answer 400 with the output in its body, and the oracle has
+  already proven execution from a value random to that probe; overwriting that
+  would turn demonstrated RCE into a false negative, which is worse than the
+  false clean this verdict removes.
+
+  **Every method is covered, including the three that decide per probe from a
+  series.** `oob`, `lookup` and `deser` take a different branch, which returned
+  before refusal was considered — so a callback run whose every probe was
+  refused reported `negative` for each of them, because no callback arrived.
+
+  Refusals are counted in **requests**, like the status tally beside them: an
+  aggregate method fires a whole ladder and reports one row, so counting rows
+  described the same run with different arithmetic and hid a partly filtered
+  series entirely.
+
+### Fixed
+
+- **The advice on a filtered run pointed at the wrong thing entirely.** "The
+  target may be patched" reads as a clean bill of health for a target that was
+  never reached; the blind-sink list names methods a filter refuses in exactly
+  the same way; and the second-order line said the target *accepted* an input
+  it had in fact rejected with a 403. None of the three fires on a refused run
+  now. It is replaced by what was actually observed, with the flags that change
+  the payload's shape.
+
+## [2.42.0] — 2026-09-23
+
+### Added
+
+- **A run that confirms nothing now names the second-order oracle.** Measured
+  against a target that stores on one endpoint and renders through a shell on
+  another — a real RCE — every probe read `negative`, and the run answered with
+  four methods that are all negative there too, because the execution does not
+  happen on the request being measured:
+
+  ```
+  --methods time   -> negative=4
+  --observe-url    -> confirmed, first run
+  ```
+
+  The one flag that works was named nowhere. It is named now, and **not gated
+  on which methods have run**: the blind-sink list is, so an operator who had
+  already tried the expensive methods — exactly the one with nothing left but
+  second order — was told only that the target might be patched.
+
+  The wording follows what the run observed rather than what it assumes. Input
+  returned verbatim means a sink that reflects without executing; input that
+  never came back means this response cannot show what became of it; and a run
+  of aggregate methods alone, which record no per-probe observation, claims
+  neither. A swallowed input is equally a blind sink or a stored one — `ping
+  <input> >/dev/null` returns nothing either — so both are named and neither is
+  picked.
+
+- **Whether the target returned the input is recorded on every probe.** It was
+  already computed on the confirmed path, where it becomes "target also
+  reflects the payload verbatim"; a negative probe never looked, and the
+  negative run is the one that has to say what it saw.
+
+  Measured against the payload that actually went out. A carrier that
+  multiplies through a filter never spells the joined `a*b` out — Liquid sends
+  `{{ a | times: b }}` and Django `{% widthratio a 1 b %}` — so an endpoint
+  echoing the whole payload recorded a measured `False`, and a target profile
+  filtering the `*` shapes leaves only those, at which point the run would
+  report "returned none of it" about a target that returned everything.
+
+  Recorded only when there was a response to look at. A delivery error is not
+  an observation, and a `False` for one would put an unmeasured claim exactly
+  where the unobserved branch belongs.
+
+## [2.41.0] — 2026-09-23
+
+### Fixed
+
+- **A candidate that confirmed RCE was never asked whether it was also a
+  deserialization or lookup sink.** The enumeration driver split the methods
+  into a cheap wave and an expensive one and skipped the expensive wave once
+  execution was proven — reasonable for `time` and `oob`, which would only put
+  a second name on one finding, and wrong for `lookup` and `deser`, which
+  report *different properties* with their own remediation. They sat on the
+  expensive side of a hand-written set of names, so the answer was never asked
+  for and never reported.
+
+  The split now comes from the tier each method declares. Anything reporting
+  `confirmed` or `needs-review` is answering *did this target execute my
+  input*; anything else is a different question and is never skipped for an
+  answer to that one.
+
+- **`--max-payloads` is spent per question, not per wave and not per
+  candidate.** Per wave it quietly doubled: a run capped at 5 sent 10 probes to
+  every candidate that did not confirm, while the cost line printed before any
+  traffic said 5. Bounding the candidate instead starves the different
+  question — the cheap methods eat the whole allowance and `deser` never runs,
+  which is the same finding lost by another route. Every method asking about
+  execution now shares one allowance, each different property gets its own, and
+  the cost line names how many questions are being asked — and **sums** the
+  estimate across them. Counting the cap once while the run grants it per
+  question advertised 44 requests for a run that sent 80, which is wrong in
+  the direction that matters for the operator bounding a monitored engagement.
+
+- **A second-order confirmation settles its carrier too.** With `--observe-url`
+  a probe can read negative in the response it drew and `confirmed` on the
+  observed channel a moment later. The stop was decided from the pre-poll
+  verdict, so the carrier kept probing after it had in fact confirmed —
+  spending the budget the stop exists to hand to carriers not yet examined,
+  which is the coverage loss this change was written to remove, reappearing on
+  the one oracle that needs a second request to answer.
+
+### Changed
+
+- **A carrier that has confirmed stops there** (`--confirm-depth first`, the
+  new default). One carrier is one method in one environment and context, and
+  once it has confirmed every further shape of it can only say the same thing
+  again.
+
+  Measured against an executing target: one candidate spent 115 of its 120
+  probes after the first confirmation, and printed 32 confirmations of which 29
+  were duplicates inside a single carrier. Those probes were not idle — they
+  were spent instead of reaching carriers never examined at all. At the same
+  budget the run went from **4 carriers examined to 23**, and from **4
+  environments reached to 9**. This buys coverage rather than saving requests.
+
+  The stop is per carrier and **never** per candidate: a candidate may confirm
+  as `unix` while a later `nodejs` carrier is the only thing a different target
+  would have shown. `--confirm-depth every` maps every shape a sink accepts,
+  which is what writing a proof of concept by hand needs, and the run reports
+  how many shapes it held back and which carriers stopped — a fourth tally
+  beside the profile drops, the safety holds and the reach notes, because it
+  says a fourth thing: the probe could have been sent and had nothing left to
+  establish.
+
+- **A method declares what one of its probes costs.** `CHEAP_DETECTION_METHODS`
+  was a set literal; it is now derived from a `costly` attribute each class sets
+  for its own reason — a real sleep, a wait for a callback, a second fetch.
+  Every hand-written list naming methods in this repository has gone stale, and
+  this one had put `lookup` and `deser` where being skipped cost findings.
+
+## [2.40.0] — 2026-09-22
+
+### Added
+
+- **A carrier may take the operands apart, and two engines need it.** Every
+  expression carrier until now substituted `__EXPR__` — the joined `a*b` —
+  which quietly assumed the engine has an arithmetic operator. Two widely
+  deployed ones do not, and both were measured as false negatives:
+
+  ```
+  liquid    every bare form missing; {{ 45013 | times: 45989 }}     -> 2070102857
+  django    every bare form missing; {% widthratio 45013 1 45989 %} -> 2070102857
+  ```
+
+  Liquid multiplies with a filter and Django with a tag, so neither form can be
+  written as a single expression — an application that really does evaluate the
+  template was reported `negative`, which is the same shape as `oob` against a
+  `${jndi:...}` sink: probes that reach the target and cannot speak its
+  language. A carrier template may now use `__A__` and `__B__` as well as
+  `__EXPR__`.
+
+  Neither payload contains the product, so a target that merely echoes the
+  payload still cannot read as `confirmed`. A test pins that for every shipped
+  carrier.
+
+### Changed
+
+- **A carrier template not parameterised by both operands is skipped, not
+  sent.** With no token at all it renders the same constant every probe; with
+  only one operand the target is never handed the other, so nothing it can
+  compute is the product RCEKit is looking for. Either way the product would
+  not be evidence the target computed anything — and a probe that cannot
+  confirm still counts toward the coverage a run reports, which is the part
+  that matters more than the wasted request.
+
+- **`--eval-engines` names every carrier the corpus ships.** Its help listed
+  three engines by hand and two were added. An operator narrowing that flag
+  reads the list and nothing else, so a stale one says an engine needs no
+  carrier when it does, and they cut the only probe that could have confirmed
+  it. A test now holds the help text to the corpus.
+
+- **What the survey measured and did not ship is recorded too.**
+  `eval_carrier_survey` in the corpus now names the engines that need no
+  carrier — nunjucks 3.2.4, tornado 6.5.10, mako 1.4.1, chameleon 4.6.0,
+  smarty 5.8.4 and Ruby's ERB all return the product from a bare form — and
+  those that are
+  out of reach. Handlebars 4.7.9 fails every bare form and, being logic-less
+  with no built-in arithmetic helper, has no template text that computes a
+  product at all. Go `text/template` 1.23 fails every bare form too, and the
+  only forms that do return the product — `{{printf "%d" <product>}}` and
+  `{{<product>}}` — hand the target the answer, so a target that merely echoed
+  them would read as `confirmed`. That is the rule a carrier lives under: **a
+  carrier may not carry its own result.**
+
+  Saying so is worth more than a carrier that cannot work, and it stops the
+  next person re-running the same survey — or shipping the Go form, which the
+  survey itself produced and which looks like a carrier until you ask what an
+  echoing target would return.
+
+## [2.39.0] — 2026-09-22
+
+### Added
+
+- **Every part of a multipart body is an injection point.** None of them were
+  before. The form branch matched on `=` appearing anywhere in the body, so a
+  `multipart/form-data` capture was split on `&` and yielded exactly one
+  candidate, named after a `Content-Disposition` line:
+
+  ```
+  form | '--X\r\nContent-Disposition: form-data; name' | body param '...'
+  ```
+
+  Every probe for that point rewrote a *part header*, so it could confirm
+  nothing, while `user`, `avatar` and `note` — the fields the form actually
+  posts — were never reached. The run still printed a point and a probe count,
+  which is the part that matters: coverage reported and not delivered reads
+  exactly like a clean target.
+
+  Several parts may post under one name — a multi-file input and a checkbox
+  array both do — so a part is addressed by its **index**, as a JSON leaf is
+  addressed by its token path. Addressing by name alone rewrote the first part
+  for every candidate, which is the same failure one level further in: three
+  files, three points, the first file probed three times and the other two
+  never touched.
+
+  `multipart` is now its own kind, recognised from `Content-Type` and decided
+  before the form branch can see the body. A file part is a candidate too — its
+  content is the value under test, while its `filename` and `Content-Type` stay
+  as captured. Verified end to end against a target that parses with the
+  standard library's MIME parser: `multipart field 'note': confirmed`, with the
+  non-vulnerable `user` field `negative` beside it.
+
+### Changed
+
+- **A GraphQL request is ordered by what can actually confirm.** Its variables
+  were already enumerated — they are JSON leaves — but they were tried in
+  document order alongside `query` and `operationName`, and those two cannot
+  confirm anything. A payload in `query` *replaces* the operation document, so
+  the server answers with a parse error before a resolver runs, and
+  `operationName` then names an operation that is no longer there. On the
+  capture this was measured against they were two points of five, each one a
+  full probe ladder.
+
+  They are moved behind the variables, **not dropped**. A server that logs the
+  query document before parsing it is reachable through exactly that field,
+  which is the route Log4Shell took through access logs, so a full run still
+  tests both and `--max-points` now cuts the least likely to pay first. A plain
+  `{"query": ...}` body with no `variables` is left in document order: it is as
+  likely to be a search API, and there the query field is the one worth testing.
+  Carrying both keys is not enough either — the `query` string has to open like
+  a GraphQL document, so `{"query": "red shoes", "variables": {...}}` keeps its
+  real injection point where a bounded `--max-points` run will still reach it.
+
+### Fixed
+
+- **The verdict table was missing a verdict.** The README said "seven verdicts
+  that are never collapsed into each other" and listed seven, while the tool
+  reports eight: `lookup-sink` was absent from the one table whose whole job is
+  to enumerate them — in a README that uses the word two tables higher, in the
+  Log4Shell CVE row, and again in the methods table. The sentence counted the
+  rows the table had rather than the verdicts there are, so the omission never
+  contradicted itself and nothing failed.
+
+  Found by reading the README end to end for this change. A test now reads the
+  table against `DETECTION_METHODS`, so a verdict a method declares and the
+  table does not carry is a failure rather than a silence.
+
+- **A multipart body now goes out with the line endings it needs.** `-r`
+  normalises the whole request to LF, so the CRLF delimiters RFC 2046 requires
+  were gone by the time anything was sent. Bodies rendered for a multipart point
+  are re-serialised canonically — probe and payload-free control alike, so the
+  two differ in the field under test and in nothing else. Part content is left
+  character for character, so a lone newline inside an uploaded text file
+  survives.
+
+## [2.38.0] — 2026-09-21
+
+### Added
+
+- **A probe shape may reach past the run's tier and be sent anyway**, when its
+  effect is one the run undoes by saying it happened. `reaches_past` is that
+  declaration, beside `safety`, which stays for an effect a notice cannot take
+  back.
+
+  Reach wins where the two pull against each other. Detection the tool could
+  have done and did not is a false negative wearing a safety label, and it
+  costs more than the noise it saves.
+
+  `deser`'s DNS gadget is the first of these, and closes an inconsistency
+  recorded a version ago: it makes the target resolve a name -- the very thing
+  `oob` and `lookup` are refused for at `safe` -- while its only gate was
+  `--oob-host`. Holding it back would have sent fewer probes at the default
+  tier. It goes, and the run reports how far it reached.
+
+  Three tallies now, because they say three different things and one number
+  would state the wrong one about all of them: the profile dropped it (it could
+  not have reached the sink), the tier held it (it could, and was not sent), or
+  it reached past the tier (it was sent, further than asked).
+
+  `lookup`'s `ldap://` and `rmi://` are the edge the rule has, and stay at
+  `stateful`: a class fetched from an address RCEKit did not choose is not
+  something a notice takes back.
+
+- **Two worked examples in the README's Quick start**, because the fullest run
+  RCEKit can make was not shown anywhere near the front.
+
+  The first is a captured request: most sinks worth testing sit behind a POST
+  with a session cookie, a content type and a body, and `--verify-url` carries
+  none of that. The second is everything the tool has -- injection-point
+  enumeration across every value in that request, every method, callbacks, the
+  top rung -- with a table of what each flag opens up and the cost line that
+  prints before it fires.
+
+  Both were run before being written down. The point the Quick start never
+  made: `--auto-params` needs `-r`, so the fullest run is not reachable from a
+  URL at all, which is worth knowing before concluding a target is clean.
+
+## [2.37.0] — 2026-09-20
+
+### Added
+
+- **A detection method declares the risk rung it needs**, and the engine reads
+  it. `SAFETY_ORDER` has labelled corpus payloads `safe` / `intrusive` /
+  `stateful` from the start and the query-language bridges followed; detection
+  methods did not. Each risky one was gated by a hand-written branch in
+  `main()` naming it, so a new method meant remembering to add another -- and a
+  probe shape with nowhere to declare its rung was deleted rather than gated.
+
+  | method | rung | |
+  |---|---|---|
+  | `reflected`, `eval`, `time`, `deser` | `safe` | compute, delay, or parse |
+  | `oob`, `lookup` | `intrusive` | makes the target open outbound connections |
+  | `file`, `write` | `stateful` | writes to the target |
+
+  `file` and `write` are gated by their own configuration rather than by the
+  rung: neither does anything until a directory to write into and a URL to read
+  it back from are named, which says more than a tier would, and asking for the
+  flag as well would refuse a command that works today. Nothing that runs today
+  stops running.
+
+- **A probe shape may need a higher rung than its method**, so coverage that
+  only makes sense at the top tier has somewhere to live instead of being
+  deleted. **`--methods lookup` sends `ldap://` and `rmi://` again**, at
+  `stateful`.
+
+  They were removed in 2.36.0 on the argument that `dns://` resolves wherever
+  `ldap://` would. That claim was too strong -- a filter catching the string
+  `dns:` and not `ldap:`, or a trimmed runtime without the DNS provider,
+  defeats it -- and a sink that takes one scheme and not the other is exactly
+  the sink this method is for. The reason they were removed was real: they
+  continue *past* resolution and connect to whatever address the answer named,
+  which is not an address RCEKit chose. That is a rung, not a reason to drop
+  coverage.
+
+- **The run says what the rung held back**, counted apart from the target
+  profile's drops and with the flag that would send it. The two say different
+  things: a profile drop means the probe *could not have* reached the sink,
+  while this means it could and the operator chose not to send it. Reporting
+  them together would state the first about the second.
+
+- **A method declares the weaker tiers it really reports**, not only its
+  ceiling. `write` reports `needs-review` for a file that is served but not
+  interpreted and `deser` for a shape fingerprint, and three separate places
+  had to know that -- the documentation tests, the benchmark's expectation
+  whitelist, and the advice printed after a clean in-band run. Each kept its
+  own answer; the docs test exempted `needs-review` for *every* method, so
+  `lookup` could name a verdict it never emits and pass.
+
+### Changed
+
+- **`--methods file` and `--methods write` with nothing configured now say
+  so by name.** They were simply not applicable before, so the run built no
+  probes and reported `nothing-tested` -- which is the quietest way this tool
+  can fail and reads much like a clean target.
+
+- **`docs/reference.md` carries a `Rung` column**, held against the class by a
+  test. The tier column already was; this is the same claim one column over.
+
+### Fixed
+
+- **A config-gated method is not re-gated by the rung at runtime.** The
+  pre-flight lets `--methods file --webroot ... --web-base-url ...` through
+  because the configuration is the gate, but the probe filter read the run's
+  default `safe` ceiling and held every probe inheriting the method's
+  `stateful` rung. The CLI accepted a documented invocation and then reported
+  `nothing-tested` -- the quietest way this tool can fail, and the thing the
+  rung work was supposed to remove rather than add.
+
+  Nothing caught it because every `file` test builds the method's config
+  directly, without `max_safety`, so the ceiling fell back to the method's own
+  rung and the probes went out. A run through the CLI with the channel
+  configured is the one thing that would have, and there is one now.
+
+- **The cost estimate applies the risk tier as well as the target profile.** It
+  counted every shape a method built, so once a rung could narrow a method the
+  pre-flight figure over-counted -- three times over for `lookup` at the
+  default tier, which is exactly the operator who narrowed the run on purpose.
+  Both paths share one predicate now, and it counts nothing, so an estimate
+  never moves the numbers the report prints.
+
+## [2.36.0] — 2026-09-19
+
+### Added
+
+- **`--methods lookup`: confirmation for an expression-lookup sink**, the shape
+  Log4Shell has, where the sink resolves a URI instead of running a command.
+
+  `oob` could not reach one. Its `applicable` does admit `java` -- a Java
+  application can shell out, so the environment is genuinely shell-capable --
+  but every probe it builds is a shell command: `nslookup`, `curl`, `certutil`,
+  `iwr`. A sink that interpolates `${jndi:...}` runs none of them, so the method
+  applied, sent its whole ladder, and came back `negative` on a target that is
+  exploitable. The README's Log4Shell row rested on the standalone listener and
+  a generated payload file, which produce no verdict row at all, so nothing in
+  the engine could reproduce that claim.
+
+  The probes are lookups and nothing else, and they depend on the injection
+  context rather than the environment, exactly as `eval`'s do. The oracle is the
+  one `oob` already uses: a token the target could only have learned by
+  resolving what it was handed. The expression resolves `<token>.<host>`, so
+  the in-process DNS listener is the entire apparatus -- no LDAP or RMI server
+  is needed, and none is started.
+
+  **It reports `lookup-sink`, never `confirmed`.** A callback proves the sink
+  resolved a URI RCEKit chose -- that it evaluated the expression it was handed.
+  It does not prove the target ran attacker code: Log4Shell becomes RCE when the
+  LDAP server answers with a loadable class. So the method gets its own
+  proven-sink tier beside `deserialization-sink`, with its own section in the
+  report, and `confirmed` keeps meaning executed.
+
+  **Only `jndi:dns://` is sent, and that is the security property rather than a
+  shortcut.** A name lookup can be nothing else. `ldap://` and `rmi://` continue
+  *past* resolution and open a connection to whatever address the answer named
+  -- by default `127.0.0.1`, which is the target's own loopback. Whatever
+  replies on :389 or :1099 is not RCEKit, so a reference could come back and a
+  class be instantiated: the tool would have crossed the line this method exists
+  to stop short of, having promised it had not. Dropping the two schemes also
+  costs no coverage -- `DnsContextFactory` ships in the JDK, so `dns://`
+  resolves wherever `ldap://` would, and on 2.15.0 it still resolves where
+  `ldap://` no longer does. The proof is the callback and the finding is "this
+  sink resolved a URI I chose" -- the same line `deser` draws, drawn here before
+  it can be crossed.
+
+  Behind the same two gates as `oob`, and sharing its listener: it needs
+  `--oob-host`, and it is held back at the default safety tier because it makes
+  the target open outbound connections. Without `--oob-host` it builds no probes
+  at all, which the engine reports as `nothing-tested` -- never `negative`. An
+  address literal is a literal whichever family it is from: `::1` and `[::1]`
+  build nothing, the same as `10.0.0.1`, because a token can only ride in a DNS
+  label. `blind_sink_advice` names the method as proving a lookup sink and not
+  execution, in a list where `oob` and `file` mean confirmed execution, and the
+  README's Log4Shell demo heading says `lookup-sink` rather than `confirmed` --
+  it was left claiming execution beside the table row that no longer does.
+
+## [2.35.5] — 2026-09-19
+
+### Fixed
+
+- **The `--insecure` notice describes the run it is in.** It was printed as
+  soon as the flag was seen, so a generation-only run, `--doctor`, or a plain
+  HTTP target all announced a TLS downgrade that never happened -- and it named
+  both downgrade rungs even on an OpenSSL build that had refused one of them.
+  Reporting a downgrade on a run that opened no TLS connection is the same
+  defect as reporting a probe that was never sent, in the one line written to
+  be an audit of the run.
+
+  Building the context and saying so are now separate questions. The context is
+  built whatever the target's scheme is -- urllib follows a redirect with the
+  handler it was given, so an `http://` target that lands on a self-signed or
+  legacy `https://` one needs it as much as a direct HTTPS target does. The
+  notice fires once, when a connection has actually reached TLS, and lists only
+  the rungs that took.
+
+- **A benchmark case may set `timeout` to zero.** `timeout or 900.0` replaced
+  an explicit `0` with the fifteen-minute default, so a case deliberately
+  bounded to no time at all ran for much longer than it asked for.
+
+## [2.35.4] — 2026-09-19
+
+### Fixed
+
+- **A wave the declared profile emptied is no longer read as a finished
+  method.** The probe filter drops probes between the method and the wire, and
+  when it emptied an adaptive method's *first* screening wave the engine took
+  the empty batch for "nothing left to send" and stopped before asking for the
+  next wave at all.
+
+  An adaptive method holds separators back in waves precisely because a filter
+  is expected: `time` screens two of them first and keeps `||`, `&&`, the
+  newline and the bare command for a second wave. A sink that strips `;` and
+  `|` removes exactly the first wave and leaves the rest intact -- and the rest
+  were never sent. Measured against a sink reachable only through `&&`, the run
+  reported **`negative`**: "the probes reached the target and found nothing",
+  about probes that were never sent. Worse than `nothing-tested`, which is at
+  least true.
+
+  The loop now tracks what the method offered separately from what the profile
+  allows to be sent, and ends only when the method itself is done. The round cap
+  still bounds it, so a method whose every wave is filtered cannot spin the
+  engine.
+
+- **The unit suite runs off Linux.** The fake vulnerable sinks in the tests are
+  POSIX command-injection points and every probe built for them is POSIX, but
+  they were executed through `os.popen` — which is `cmd.exe` on Windows. The
+  sink the test says executes did not execute, so the oracle correctly reported
+  no execution and five tests failed for a reason unrelated to the code under
+  test. Worse, the tests asserting a *negative* stayed green throughout: a
+  broken fixture that keeps its controls passing is the failure this project
+  takes seriously everywhere else. The sinks now name a POSIX `sh` explicitly;
+  on Linux and macOS that is the `/bin/sh` they always used.
+
+- **A deliberately dead target no longer costs minutes.** A closed loopback port
+  answers with a RST on Linux and is silently dropped on Windows, where each
+  probe waits out the SYN retry instead — measured at ~2s per probe. The
+  benchmark harness's own unreachable-target case paid that for the full ladder
+  twice, once for the vulnerable half and once for the control: 1800s for one
+  test, longer than the other 534 together. Bounded to three probes it measures
+  9.1s, and three probes prove "nothing reached the target" exactly as well as
+  forty do.
+
+- **A captured-request fixture reaches disk byte for byte.** The cleartext-capture
+  tests write a raw HTTP request whose text already spells its own CRLF line
+  endings, through `Path.write_text` — which on Windows translates the newline of
+  each one again. The file on disk held a doubled carriage return, the parser
+  found no headers, and two tests failed against a fixture that had stopped being
+  an HTTP request at all.
+
+  The third test in that class passed throughout, for the wrong reason: it
+  asserts that a notice is *absent*, and a request that cannot be built prints no
+  notice either. A control that stays green while its fixture rots is precisely
+  what this project refuses to accept from a benchmark case, so the fixture now
+  has a guard of its own. The capture is written as bytes; the `newline` argument
+  that would say the same thing arrived in Python 3.10, and this project
+  supports 3.8.
+
+## [2.35.3] — 2026-09-18
+
+The target profile an operator declares now reaches the probe ladder, not just
+the corpus.
+
+### Fixed
+
+- **`--deny-chars` / `--max-length` reach the detection probes.** They were
+  applied by `_filter_by_profile`, which drops *corpus records* — and stopped
+  there. The probes a detection method builds from those records went out
+  regardless, so a run that had been told "this target strips quotes" still paid
+  for every quote-carrying rung of the ladder, on requests structurally unable
+  to confirm. Those requests are not free: on a captured request with `-p all`
+  they are the budget the next injection point never got. The filter now sits at
+  the engine, where every probe passes through it — deliberately not inside
+  `_wrap_variants`, because `_space_free_probes`, the query-language bridges,
+  `eval`, `oob` and `deser` each build payloads without going through that
+  helper, and a gate that reaches some methods and not others is the side path
+  that once left `file`/`time`/`oob` unable to send the raw rung.
+
+  Denying a character narrows the ladder rather than emptying it: a target that
+  strips `;` is still probed through `|`, `||`, `&&` and the newline, which is
+  what the separator table has always been for.
+
+  Checked on the literal payload, before the delivery layer percent-encodes it
+  for its injection point. That is stricter than the corpus check, which is
+  applied to the encoded payload and so lets a URL-encoded quote through a quote
+  filter. The layers genuinely differ: transport encoding is undone by the
+  server before the value reaches the sink, so a percent-encoded quote is still
+  a quote when the application's own filter sees it.
+
+- **A profile strict enough to remove every probe reports `nothing-tested`.**
+  Not `negative`, which would read as "not vulnerable" from a run that sent
+  nothing. The message names the profile as the cause and the characters a probe
+  would have to avoid, instead of the generic advice to widen `--environments` —
+  which is not what emptied the run.
+
+- **The cost estimate follows the profile.** `[detect] cost:` builds the probes
+  and counts them, so it now counts the ones that will actually be sent. An
+  estimate that ignores a filter is wrong precisely for the operator who
+  narrowed the run on purpose.
+
+### Changed
+
+- A run that dropped probes says so, with the reason and a count per reason. A
+  ladder that shrinks quietly is the one way this filter could manufacture a
+  false negative, so the removal is stated rather than left to be inferred from
+  the traffic.
+
+## [2.35.2] — 2026-09-18
+
+### Fixed
+
+- **`--insecure` now reaches a legacy TLS stack, not just an untrusted one.**
+  Turning certificate verification off is not the same as completing a
+  handshake. OpenSSL 3.x ships security level 2, which refuses the key sizes and
+  signature algorithms that software of the era this tool gets pointed at still
+  offers — Webmin 1.910, the build the README's `reflected` row rests on,
+  answers a default client with `SSLV3_ALERT_HANDSHAKE_FAILURE` and nothing
+  else. Every probe then came back `error`: correct, and useless. The run was
+  honest about having measured nothing, and the sink behind that handshake was
+  never tested at all. Found by running the coverage benchmark, which failed at
+  its readiness gate against a container that was up and answering.
+
+  `--insecure` now also lowers the security level and the minimum protocol
+  version. This does not widen exposure: with `check_hostname = False` and
+  `CERT_NONE` the connection is already unauthenticated, so an active attacker
+  is already unconstrained — accepting a 1024-bit key or a SHA-1 signature on
+  top of that gives away nothing that was still being held. What it buys is the
+  difference between testing the target and reporting that it could not be
+  reached. A run without the flag is untouched and still verifies certificates.
+
+- **The benchmark's readiness gate is as permissive as the tool it gates.**
+  `wait_for_target` built its own strict context, so a case against deliberately
+  old software reported "target never became ready" about a container that was
+  up — a case failure with nothing wrong in it.
+
+### Changed
+
+- A run that passes `--insecure` states the full extent of the downgrade on its
+  first line. The flag gives up more than certificate identity now, and an
+  operator on a monitored engagement should read that in the transcript rather
+  than infer it from the help text.
+
+## [2.35.1] — 2026-08-21
+
+A robustness pass over error handling: no new capability, four ways the tool
+could crash or mislead on input it did not choose.
+
+### Fixed
+
+- **A truncated error response no longer ends the run.** Reading an
+  `HTTPError`'s body happens *inside* the `except` handler, where the sibling
+  `except Exception` cannot reach it — so a target that promised a
+  `Content-Length` it never delivered raised `ConnectionResetError` straight out
+  of `main()`, taking every probe already fired with it. The read is now
+  guarded: the status still comes back, an unreadable body is reported empty.
+  A body that *does* arrive is still returned in full — the 500-stack-trace
+  confirmations that branch exists for are unaffected.
+- **A `--target-profile` is checked before it is used.** A profile is written by
+  hand, so a typo in one is ordinary; it surfaced as a traceback. A top level
+  that is not a JSON object, `deny_chars` that is not text, a `max_length` that
+  is not a number, a selector field that is not a list of names — each is now an
+  operator-readable `[!]` message and exit 1, the same way the sink-shape fields
+  already behaved. Twelve inputs that produced an `AttributeError` or a
+  `TypeError` now produce a sentence.
+- **A selector field given as one string means one name.** `"environments":
+  "unix"` in a profile was iterated character by character, matched nothing, and
+  the empty run that followed was reported as a success. A string is now split
+  on commas: `"unix"` is `["unix"]`, `"raw, html"` is `["raw", "html"]`.
+- **Unknown `--environments` and `--encodings` are named.** Both were silent, so
+  `--environments linux` — the corpus calls it `unix` — produced an empty file
+  and exit 0, indistinguishable from a target with no payloads for it. Both now
+  warn and list the names that exist, as unknown contexts and categories already
+  did. An empty result is also no longer announced as "Successfully generated 0
+  payloads"; it says the selection matched nothing and points at the filters.
+- **A callback cannot rewrite the operator's terminal.** The host and path of an
+  OOB callback are chosen by the target. Printed raw, an ESC byte let that
+  target colour, erase and rewrite lines — hiding a genuine `[HIT]` behind
+  `\x1b[2K\r`, or forging one that never arrived. Control characters are now
+  escaped for display as `\xNN`; the recorded hit and the `--listen-log` JSONL
+  keep the bytes verbatim.
+
+## [2.35.0] — 2026-08-20
+
+### Added
+
+- **RCEKit is installable from PyPI**: `pipx install rcekit` (or
+  `pip install rcekit`) puts an `rcekit` command on PATH. Published through PyPI
+  Trusted Publishing from a GitHub release — no API token, no repository secret.
+
+  It ships as a **single-module distribution**, not a package tree. `rcekit.py`
+  stays one file at the repo root and still runs alone from a `curl` on a jump
+  box or an air-gapped host; installing is a second supported shape, not a
+  replacement for the first.
+
+### Changed
+
+- **`import rcekit` no longer has side effects.** Logging was configured at
+  module scope, and `logging.FileHandler` opens its file when it is constructed,
+  so merely importing the module wrote `rcekit.log` into whatever directory the
+  interpreter happened to be in. Handler setup moved into `configure_logging()`,
+  called from `main()`. Running the CLI still writes `rcekit.log` exactly as
+  before.
+
+- **`main()` takes an optional `argv`** and returns an explicit `int`, so the
+  console-script entry point is a plain zero-argument call and tests can drive
+  the CLI in-process. Every exit code is unchanged.
+
+- **The built-in corpus is no longer reported as a missing file.** With nothing
+  but `rcekit.py` — an installed wheel, or the single-file copy — there is no
+  `templates/` directory, the embedded corpus *is* the corpus, and the run is
+  now silent about it; `--doctor` names it `built-in (embedded in rcekit.py)`
+  and reports OK. A `templates/` directory that exists *without* its
+  `payloads.json` still prints the notice, because that one is a real finding.
+  A corpus that is present but corrupt, and an explicit `--template-file` that
+  is missing or corrupt, still refuse to run and exit non-zero.
+
+- The log file handler now runs at `DEBUG` while the console stays at `INFO`, so
+  detail worth having when reconstructing a run no longer lands on the
+  operator's terminal.
+
+## [2.34.1] — 2026-08-17
+
+Documentation only; no behaviour change.
+
+### Changed
+
+- **New tagline: "`confirmed` means the target executed the input. `negative`
+  means the probes reached it."** The old one — "prove RCE, don't guess it" —
+  claimed the tool always proves. It does not, and does not need to: the value
+  is that each verdict has a mechanical meaning, in both directions. A promise
+  can be broken; a definition cannot.
+
+  Both halves name their actor and object on purpose. "Executed" alone reads as
+  though *RCEKit* executed something; the claim is about the target. And
+  `confirmed`/`negative` are the verdict values as the code spells them, not
+  looser words like "clean".
+
+- **The README caught up with the engine.** It still described "two verdict
+  tiers" when there are seven, and the comparison table predated the last five
+  releases. Rewritten around what a verdict asserts, with the tier table as the
+  centrepiece and a section on the half no other tool has: `error` and
+  `nothing-tested` exist so a run that tested nothing is never reported as
+  clean.
+
+  The comparison table gains the classes added since it was written — Windows
+  `cmd.exe`/PowerShell sinks, upload → write-then-execute, second-order
+  execution, query-language bridges and deserialization sinks — and the "reach
+  for something else" note now says plainly that sqlmap owns the database and
+  RCEKit's bridges only prove the OS is reachable from a text parameter.
+
+### Fixed
+
+- Two claims in the engagement-controls table were wrong and are now accurate:
+  the observed-channel fetch sends **no** credentials unless given a request with
+  `--observe-request` (only the `file` read-back inherits them, same-origin), and
+  an unanswered `--observe-url` is a warning about a partly blinded run rather
+  than a `nothing-tested` verdict.
+- The "mechanisms that produce `inconclusive`" list said four and listed five.
+
+## [2.34.0] — 2026-08-17
+
+Deserialization **sink** detection, and a verdict that is deliberately not RCE.
+
+Deserialization RCE (fastjson, shiro, weblogic, jenkins) cannot be confirmed by
+the value-oracle model: the payload is a serialized object graph and gadget
+selection is classpath-specific, so whether execution is reachable depends on
+jars RCEKit cannot see. That stays out of scope. The honest middle step is
+showing the endpoint parses the data at all — a real finding, and the
+prerequisite for every gadget chain.
+
+### Added
+
+- **`--methods deser`**, which **never emits `confirmed`**. Its strongest
+  outcome is a new verdict, `deserialization-sink`, reported in its own section
+  that states outright that reaching RCE from there depends on classpath
+  gadgets. Collapsing it into `confirmed` would break the one guarantee the tool
+  rests on; collapsing it into `needs-review` would throw away a proven finding.
+
+  `deserialization-sink` sits below both RCE tiers in the collapsed verdict: it
+  is proven, but a *suspected* RCE outranks a proven non-RCE in triage.
+
+- **Two oracles of deliberately different strength.** `shape` (no listener
+  needed) sends a well-formed object stream, the same stream truncated, and the
+  format's magic bytes plus random noise of the same length, and asks whether
+  the endpoint answers the well-formed one differently from both — a
+  fingerprint, so `needs-review` only, never promoted. `dns` (needs
+  `--oob-host`) sends a gadget whose only side effect is a name lookup.
+
+- **A URLDNS builder for Java serialization.** A `HashMap` holding one
+  `java.net.URL`: `HashMap.readObject` hashes the key, `URL.hashCode` asks for
+  the host address, the JVM resolves the name. It references no class outside
+  `java.util`/`java.net`, so there is nothing in it that can run — the callback
+  proves the object graph was reconstructed and no more.
+
+  Built in Python rather than declared in the corpus because the URL host is
+  length-prefixed *inside* the stream and changes per probe. Its constant parts
+  are the exact bytes OpenJDK's own `ObjectOutputStream` emits for that graph,
+  and the result was **verified against OpenJDK 21**: it deserializes to
+  `HashMap{http://<host>/=rk}` and issues a DNS query for `<host>`, with no code
+  execution.
+
+- **A `deser_probes` corpus section** with `java`, `php`, `dotnet`,
+  `python_pickle` and `fastjson`, plus `--deser-formats` to narrow it. Only
+  `java` and `fastjson` carry a DNS gadget: PHP and .NET chains all run through
+  magic methods or type confusion, so there is no honest DNS-only probe for them
+  and they get the shape oracle alone.
+
+- Response signatures for the shape differential drop long digit and hex runs,
+  so request ids and timestamps on an otherwise identical error page do not make
+  every endpoint fingerprint as a parser.
+
+### Changed
+
+- The README's scope note now says precisely what changed and what did not:
+  deserialization **gadget chains** remain out of scope, while the **sink** is
+  now reported in its own tier.
+
+## [2.33.0] — 2026-08-17
+
+Query-language bridges. Several RCEs pass through a query language before
+reaching the OS — Postgres `COPY … FROM PROGRAM`, MSSQL `xp_cmdshell`, XXE
+`expect://` — and the injection point is an ordinary text value, so the oracle
+model already fitted. Only the carriers were missing.
+
+### Added
+
+- **A `bridges` section in the corpus**, declared like `eval_carriers` so
+  coverage grows without touching Python. Each bridge names the shell it
+  reaches, its safety tier, its prerequisites and, where it creates something,
+  the statement that removes it: `postgres_copy_program` (`/bin/sh`,
+  `stateful`), `mssql_xp_cmdshell` (`cmd.exe`, `intrusive`) and `xxe_expect`
+  (`/bin/sh`, `intrusive`).
+
+- **`--bridges none|auto|NAMES`** rides the command probes through them. A
+  bridge is a **carrier, not an oracle**: it wraps the command `reflected`,
+  `time` and `oob` already build, so those methods prove execution through it
+  and inherit every tier guarantee rather than re-deriving one. Off by default,
+  because a bridge payload is SQL or XML syntax and on an ordinary shell sink it
+  is a request that cannot confirm.
+
+  Three properties follow from that framing. A bridge only gets a core written
+  in its own dialect — `xp_cmdshell` hands its argument to `cmd.exe`, so pairing
+  it with a POSIX `$((a+b))` would send inert text. No separator is prepended:
+  inside `COPY … FROM PROGRAM '…'` there is no running command to break out of.
+  And the record's context still applies, so `--contexts sql` and a bridge
+  compose instead of each reinventing the other.
+
+- **The safety ordering governs bridges** exactly as it governs every corpus
+  payload: a `stateful` bridge needs `--verify-active-risk stateful`, and the
+  pre-flight names the tier each held-back bridge actually requires rather than
+  sending the operator to raise the ceiling further than the run needs.
+
+### Changed
+
+- **An aggregate method's result now carries its cleanup line.** `time` reports
+  one row for a whole probe series, so a stateful bridge on the one oracle that
+  reliably proves a query-language sink was the one that never said how to clean
+  up after itself.
+
+### Not built, deliberately
+
+- **MySQL UDF execution** is a multi-stage chain — write a shared object into the
+  plugin directory, then `CREATE FUNCTION` — not something a single probe can
+  carry. There is no stub for it.
+- **MongoDB `$where`** is a boolean-only channel (its JS sandbox cannot reach a
+  shell), so it needs a different oracle rather than this one.
+  `mongo-express/CVE-2019-10758` is a plain JS `eval` sink that `--methods eval`
+  already covers.
+
+The three shipped bridges are documented syntax but **not validated against live
+databases here** — this build environment has no container runtime. Each corpus
+entry says so in its `verified` field rather than implying a test that did not
+happen.
+
+## [2.32.0] — 2026-08-17
+
+The second-order oracle. Execution frequently happens on a **different request**
+than injection — stored SSTI rendered on a profile page, a payload written to a
+log a template engine later renders, a queued job run asynchronously. The engine
+diffs the response it injected into, so every one of those read `negative`
+however exploitable the target was.
+
+### Added
+
+- **`--observe-url URL`** names the endpoint where the execution surfaces. It is
+  read after each probe and then polled after the batch, and a probe whose
+  computed value turns up there is upgraded to `confirmed`.
+
+  It stays fully differential, which is why it reaches `confirmed` rather than
+  `needs-review`: the value was computed locally from operands random to that
+  probe, it must be absent from a snapshot of the endpoint taken **before any
+  probe was sent**, and — the rule that carries the weight — a probe's value is
+  looked for there **only when the probe's own payload does not contain it**.
+
+  Without that last rule the oracle would be a false-positive generator: `file`
+  and `oob` expect a random token that sits verbatim in the payload, so a target
+  that merely stores the payload and renders it back would hand that token
+  straight to the observed page and every such probe would confirm without
+  executing anything. Measured against a store-and-echo target: **0**
+  confirmations. The computed-value methods pass the same rule for the opposite
+  reason — reflection returns `$((a+b))`, never the sum — so it selects them
+  without naming them, and a method added later inherits the right answer.
+
+- **`--observe-request FILE`** takes a captured request instead, for the common
+  case where the page a stored payload renders on is behind a login. It needs no
+  `FUZZ` marker: the observed endpoint is read, never injected into.
+
+- **`--observe-poll` / `--observe-timeout`** control the polling window
+  (defaults 5s and 60s). One poll always happens, even at a zero timeout.
+
+- Every probe result carries an `observe_status` in `--detect-json`:
+  `confirmed`, `polled` (read, value not there), `in-control`, `not-observed`
+  (not eligible) or `unreachable`. When the endpoint never answered, the run
+  says so outright — negatives decided without ever reading the observed channel
+  are not second-order negatives.
+
+### Changed
+
+- The observed channel is read once after **each** probe as well as polled after
+  the batch, so a run with `--observe-url` sends roughly twice the requests.
+  Batch-then-poll alone is only correct for a channel that *accumulates* (a log,
+  a comment list); where the store overwrites — a profile field, which is the
+  shape this oracle most exists for — every probe but the last is gone by the
+  time the batch poll runs, and the oracle confirmed nothing. The extra read is
+  skipped for probes that are already confirmed in-band or not eligible, so
+  `file` and `oob` add none.
+
+Observing is additive throughout: the in-band verdict is computed exactly as
+before and only a non-`confirmed` one can be upgraded, so a run without the flag
+is byte-for-byte unchanged and a run with it can only gain findings.
+
+## [2.31.0] — 2026-08-17
+
+The `write` method: a write primitive proven to be RCE by executing what it
+wrote. A whole family of targets was invisible — `tomcat/CVE-2017-12615` (PUT a
+JSP), `activemq/CVE-2016-3088`, `weblogic/CVE-2018-2894` — because the vulnerable
+request *stores a file* rather than evaluating anything. Nothing is computed in
+its response, so `reflected` and `eval` correctly returned `negative` on targets
+that are fully exploitable.
+
+### Added
+
+- **`--methods write`** — the inverse of `file`. `file` assumes execution exists
+  and uses a write as proof of it; `write` assumes a write primitive exists and
+  uses execution of the written file as proof of RCE. The probe is the file's
+  *content*: a one-liner computing a product on random operands, delivered
+  through the ordinary injection point.
+
+  The fetched file is read in three tiers, and the middle one is the reason the
+  method exists:
+
+  | fetched file contains | verdict | means |
+  |---|---|---|
+  | the product | `confirmed` | written **and** executed |
+  | the one-liner, verbatim | `needs-review` | arbitrary file write, not interpreted |
+  | neither | `negative` | no write, or not served there |
+
+  An upload directory that is served but not interpreted is a real finding and
+  is not remote code execution, so the tiers are never merged.
+
+- **`--write-url-template URL`** names where the stored file is served — the
+  channel the proof comes back on, and the flag the method is gated on.
+
+- **`--write-lang`** picks the file types: `auto` (default) reads the extension
+  off the read-back URL, or name any of `jsp`, `jspx`, `php`, `aspx`, `erb`.
+  `jsp`/`aspx`/`erb` share the `<%= %>` delimiters, so their probes are
+  byte-identical and cost one request between them; with no extension to read,
+  `auto` writes all five in three requests.
+
+### Changed
+
+- **A `needs-review` finding now prints its cleanup line too.** It used to
+  appear only under `confirmed`, which was already thin and is wrong for this
+  method: a `write` reaching `needs-review` means the file *is* on the target,
+  just not interpreted, so the artifact would have been left there unmentioned.
+
+- The write method's operands are drawn once per run rather than once per
+  carrier, so the file is written once instead of once for each of the ~13
+  `(environment, context)` carriers. For a state-changing method that is not a
+  request-count saving, it is a blast radius. Still fresh per run, which is what
+  makes the product unforgeable.
+
+- The write method declines the break-out contexts (`sql`, `javascript`,
+  `shell_*`, …) and keeps the transport ones. Its payload is a whole file body:
+  there is nothing to break out of, and wrapping it in `'; … -- ` would write a
+  broken file. A run narrowed past `raw` and the transport contexts is told so
+  rather than reporting a clean negative.
+
+## [2.30.0] — 2026-08-17
+
+Per-dialect shell probes. `$((a+b))`, `sleep` and `$(echo TAG)` are POSIX
+constructs: on a cmd.exe or PowerShell sink they are inert literal text. The
+dialect was inferred from the corpus environment alone, so a run could send a
+probe no shell on the target would ever execute — including on the carrier whose
+context is literally named `powershell`.
+
+### Added
+
+- **`--sink-env auto|unix|windows|powershell`** states which shell runs the
+  injected command. The computed-value core, the separators and the break-out
+  contexts are all chosen from it. `auto` (the default) infers it per carrier;
+  pin it when the corpus environment names the *application runtime* rather than
+  the OS — `--environments php --sink-env windows` is a PHP application on IIS,
+  which no inference can see.
+
+- **A PowerShell probe shape for every shell method**, validated against
+  pwsh 7.4: `Write-Output T1$(a*b)T2` for `reflected` (an unquoted argument is an
+  expandable string, so the core carries no quote and the quote-wrapping
+  contexts can still carry it), `Start-Sleep -Milliseconds N` for `time`,
+  `Set-Content` for `file` and `iwr -useb` for `oob`. PowerShell was previously
+  reachable by no probe in any method.
+
+- **cmd.exe and PowerShell carriers for the `dotnet` environment.** It is the
+  one corpus environment that names a platform, and it was taking the POSIX
+  shape — so .NET on Windows, the case the environment exists for, was the case
+  it could not confirm on. Every other runtime keeps the POSIX shape: a language
+  does not say which OS it runs on.
+
+### Fixed
+
+- **The `powershell` carrier was written in cmd.exe.** Every `windows` carrier
+  took the `for /f ... ('set /a a+b')` core regardless of context, so the one
+  carrier explicitly shaped for PowerShell sent a payload PowerShell cannot
+  execute. The dialect now follows the carrier's context first, and a carrier's
+  break-out variants stay in its dialect rather than re-deriving from the
+  environment.
+
+- **`Set-Content`, not `>`, for the PowerShell write.** In Windows PowerShell
+  5.1 the redirect is `Out-File`, whose default encoding is UTF-16LE: the write
+  lands and the read-back still does not find the token, so the probe reports
+  negative on a target it owns.
+
+### Changed
+
+- **cmd.exe no longer gets the `sq`, `dq` and `subshell` carriers.** It has
+  neither a comment character to swallow the sink's tail nor a
+  command-substitution syntax, so those four carriers per Windows run were
+  requests that could only come back negative. PowerShell takes the quote
+  break-outs and `$( )` — both measured — but not the backtick, which is its
+  escape character rather than a substitution.
+
+- **PowerShell's separator sweep carries no pipe.** `cmd | Start-Sleep
+  -Milliseconds 500` is a parameter-binding error, not a fresh command with
+  stdin attached the way a POSIX pipe is, and it fails that way for every cmdlet
+  the probes use. `;`, a newline and (on PowerShell 7) `&&`/`||` remain.
+
+- The pre-flight plan prints the sink shell alongside the sink shapes, and a
+  pinned dialect narrows the printed ladder to the rungs it has syntax for.
+
+## [2.29.0] — 2026-08-17
+
+Generalised read-back for the `file` method. It required a writable **web root**
+the tester already knew, which ruled out every other way a target can hand a
+file back — on exactly the internal, no-egress targets the method exists for.
+
+### Added
+
+- **`--file-write-path DIR` + `--file-read-url URL`** name the two halves of the
+  read-back channel directly, so an LFI endpoint, a download or export handler,
+  an attachment fetcher or a `/tmp`-backed preview all work. The template takes
+  `{name}` (the filename), `{path}` (the full server-side path) and `{path_enc}`
+  (that path percent-encoded); only those three are substituted, so a URL that
+  legitimately contains braces survives unchanged.
+
+  Measured against a target with a download handler and nothing serving the
+  write directory: the web-root form confirms **0** — reporting an exploitable
+  target clean — and the general form confirms **7**.
+
+### Fixed
+
+- **The read-back fetch now carries the run's headers**, so an authenticated
+  download, export, attachment or LFI handler can actually be read. It went out
+  bare, which barely mattered while the channel had to be a web root — static
+  file serving is rarely authenticated — and became the likely case the moment
+  the channel could be an application endpoint. Measured against a handler
+  behind a bearer token: the write executed on every probe and the verdict was
+  `negative`, "token absent from the fetched file". Now 7 confirmations on the
+  same target.
+- **Credentials are carried only to the same origin.** A read-back URL on
+  another host is someone else's server, and replaying the target's session
+  cookie or bearer token to it would leak the credential, so those headers are
+  dropped there while the rest still go — and the run says so, because the
+  symptom would otherwise look like a clean target. `Content-Type` and
+  `Content-Length` are dropped from the fetch too: they describe a body the GET
+  does not have.
+
+### Changed
+
+- **`--webroot` / `--web-base-url` are now the web-root alias** for the general
+  form: a web root is just the case where the read URL is the base plus the
+  filename. Existing command lines are unaffected. Both are resolved in one
+  place inside the method, so the alias and the general form cannot drift — and
+  the gate, the pre-flight banner and the blind-sink advice all ask that same
+  resolver instead of testing for the webroot pair.
+- `blind_sink_advice` reads its flags defensively, so an args-like object
+  missing a newer field costs a line of advice rather than a traceback.
+
+## [2.28.0] — 2026-08-17
+
+Injection-point enumeration. `-p NAME` needed the tester to already know which
+parameter was the sink, so a capture's other candidates — including the headers
+and nested JSON leaves that carry some of the highest-value classes — were never
+tried.
+
+### Added
+
+- **`-p all` / `--auto-params KINDS`** expands one captured request into every
+  candidate injection point and runs the selected `--methods` against each.
+  Query values, JSON leaves addressed by path (`user.profile.name`, `tags[1]`),
+  form fields, cookie crumbs and headers, each rewritten in **its own**
+  serialization rather than blanket-encoded. Verified end to end: a sink
+  reachable only through `User-Agent` is confirmed from `-r request.txt -p all`
+  with no manual header selection.
+- **`--point-order fast|thorough`** — `fast` tries a curated high-yield header
+  list (the headers real published RCEs inject through); `thorough` adds every
+  remaining non-hop-by-hop header. **`--max-points N`** bounds the run and
+  reports what it dropped. **`--include-path-segments`** is opt-in, because
+  rewriting a path segment usually just produces a 404.
+- **The run states its cost before sending it** —
+  `6 points x ~61 probes = at least 372 requests` — via a new
+  `estimate_detection_probes`, which builds the probes and counts them without
+  firing any. Enumeration multiplies an already-laddered probe count by the
+  candidate count, and an operator on a monitored engagement has to see that
+  before it happens rather than infer it from the traffic.
+- **Findings name the point they came from**: `[reflected/unix/raw] at header
+  'User-Agent' ...`.
+
+### Changed
+
+- **Each candidate carries its own payload-free control.** Differencing a header
+  probe against a query probe's control would compare two different responses
+  and prove nothing.
+- **Cheap methods run first per candidate, and a candidate stops at its first
+  confirmation.** `reflected` and `eval` cost one response each; `time` sleeps
+  and `oob` waits for a callback, and on a candidate that has already proven
+  execution those buy a second name for the same finding. Candidates that stay
+  clean still get every method, and single-point runs are unchanged.
+- A JSON leaf is **replaced, never created**. Assigning to a missing key would
+  have injected into a field the application never sends — a probe that cannot
+  say anything about the parameter that does exist. Caught by its own test.
+- **JSON points are addressed by tokens, not by a joined path string.** A key may
+  itself contain the separator: `{"user.name": ..., "user": {"name": ...}}`
+  rendered *both* leaves as `user.name`, so the literal key was never probed and
+  both candidates mutated the nested field — a false negative and a misattributed
+  finding at once. Tokens remove the ambiguity, and the display form
+  bracket-quotes such a key (`["user.name"]`) so the two stay distinguishable on
+  screen.
+- **A deeply nested captured body no longer ends `-p all` with a traceback.**
+  `json.loads` recurses in C, so `RecursionError` joins the caught exceptions in
+  both the enumerator and the placer, as it already had in the response-channel
+  parser. The body yields no candidates; the rest of the request still enumerates.
+- **The cost estimate honours `--max-payloads`.** It counted every probe the
+  carriers could produce while the run stops at the cap, so the figure was wrong
+  exactly when the operator had reached for the budget guard.
+- `Host`, `Content-Length`, `Cookie` and the hop-by-hop headers are never
+  candidates: injecting into those changes the request's plumbing rather than
+  testing the application, and two of them are rebuilt by the delivery layer.
+
+## [2.27.0] — 2026-08-17
+
+Engine carriers for the `eval` probe. Three template engines evaluate the
+injected expression perfectly and still made RCEKit report `negative`, because
+what came back was not the bare product the oracle searches for.
+
+### Added
+
+- **`eval_carriers` in the corpus**, and `--eval-engines auto|<names>` to select
+  them. A carrier wraps the same random-operand arithmetic in an engine-specific
+  form; it never changes the oracle, and the bare probes still run first. Each
+  entry records `notes` (why it exists) and `verified` (what it was measured
+  against). Declarative, so a new carrier is a JSON entry rather than a code
+  change.
+
+  | Engine | Bare `${a*b}` returned | Carrier | Carrier returned |
+  |---|---|---|---|
+  | Freemarker | `2,070,761,401` (locale grouping) | `${(a*b)?c}` | `2070761401` |
+  | Velocity | `${a*b}` verbatim — a *reference*, not an expression | `#set($rk=a*b)$rk` | `2070761401` |
+  | Thymeleaf | `${a*b}` verbatim — needs inlining brackets | `[[${a*b}]]` | `2070761401` |
+
+  Measured against freemarker 2.3.32, velocity-engine-core 2.3 and thymeleaf
+  3.1.2, running RCEKit's own generated probes through each engine: bare form
+  `CONFIRMS=no`, carrier `CONFIRMS=YES`, for all three.
+- **The evidence line names the carrier** — `target computed '3979016' via the
+  freemarker carrier` — so a finding says which engine quirk it worked around.
+  A bare confirmation reads exactly as before.
+
+### Notes
+
+- **Carriers are not sandbox escapes, and no sandbox-escape carrier ships.** The
+  premise that a sandboxed engine blocks the arithmetic probe did not survive
+  measurement: a member-access sandbox restricts method and field access, and
+  arithmetic needs neither. With OGNL member access denied for *everything*,
+  `40277*51413` still returned `2070761401` while `@java.lang.Math@max(1,2)` was
+  blocked; SpEL's restricted `SimpleEvaluationContext` and Jinja2's
+  `SandboxedEnvironment` behaved the same way. The bare probes already cover
+  those engines.
+- The frequently-cited OGNL escape `(#_memberAccess=@ognl.OgnlContext@DEFAULT_MEMBER_ACCESS)`
+  additionally targets a field that **no longer exists in OGNL 3.3.4**, so on a
+  current engine it is a probe that can only come back negative.
+
+## [2.26.0] — 2026-08-16
+
+The sink-shape ladder. An injected value lands in a *shape* — mid-command,
+inside quotes, as the whole command — and the shape decides what can reach it.
+Two shapes had no probe that fitted, so a genuinely exploitable target reported
+clean.
+
+### Added
+
+- **`--sink-shape auto|sep|raw|chain|newline|dq|sq|subshell`** (comma-separated)
+  names which shapes the shell probes try. `auto` is the whole ladder and the
+  default. Underneath it selects the existing separator sweep and break-out
+  contexts, so naming a rung narrows a supported run rather than switching on a
+  parallel path. The plan is printed before anything is sent, because the ladder
+  multiplies request count and an operator on a monitored engagement needs to
+  see the cost first.
+- **The `subshell` rung — `$(...)` and backticks.** Reaches a value sitting
+  inside double quotes *without closing the quote*, which is the one case a
+  quoted break-out loses to a filter on the quote character itself. Measured
+  against `system("echo PING \"$input\"")`: with `"` stripped, `dq` is inert and
+  both substitution forms execute; with `$` stripped, `dq` executes and the
+  backtick form still does. Both ship because they survive different filters.
+
+  **Which method it helps is the counter-intuitive part.** `reflected`'s core is
+  `$((a+b))`, which the shell expands inside double quotes anyway, so that
+  method already confirmed there. The methods whose core must actually *run* —
+  `time` (a sleep), `file` (a redirect), `oob` (a fetch) — are completely inert
+  inside those quotes. On a quote-filtering sink, `--methods file` went from **0
+  confirmations to 2**: it had been reporting an exploitable target as clean.
+- **The `raw` rung is now part of `auto`, for every shell method.** A
+  `qx/$input/`-style sink, where the input is the whole command, previously
+  needed `--sink-raw` — so it reported clean unless the operator already
+  suspected the shape. One extra probe per carrier buys it. `--sink-raw` keeps
+  its meaning as the narrowing alias for `--sink-shape raw`, and no existing
+  command line changes behaviour. `reflected`, `file`, `time` and `oob` all
+  build their candidates through one `_separator_candidates` helper, so a rung
+  cannot reach some methods and not others; `time` screens it in its second
+  wave, alongside the separators it holds back.
+
+### Fixed
+
+- **A method that builds no probes no longer reports `negative`.** An aggregate
+  method asked to judge zero samples answers honestly — "no delay was observed",
+  "no callback arrived" — and that reads as "not vulnerable" from a run that
+  tested nothing. The engine now emits no row for a carrier that produced no
+  probes, which lets its own loud nothing-tested path fire instead. Reachable
+  through any narrowing that leaves a carrier with nothing to send.
+
+### Changed
+
+- **The pre-flight sink-shape plan is computed from the effective run**, not
+  from the `--sink-shape` value. `--separators`, `--contexts` and `--sink-raw`
+  each narrow the ladder, so printing the flag described a run that would not
+  happen — and this output is presented as an audit of the traffic about to be
+  sent. `effective_sink_shapes` is the single source of truth the engine and the
+  plan both read.
+- **The backtick context drops probe shapes that carry their own backtick.**
+  Backticks do not nest, so such a probe closes the outer substitution early and
+  could only ever come back negative. `$( )` does nest and keeps every shape.
+- **Naming `--separators` now implies the sink is separator-led**, so the `raw`
+  rung is dropped unless `--sink-shape` names it explicitly. A profile with
+  `sink_needs_separator` drops it for the same reason. Both keep an explicitly
+  narrowed run from being widened behind the operator's back.
+
+## [2.25.0] — 2026-08-16
+
+A coverage benchmark, so a claim about what RCEKit confirms can be checked
+instead of asserted. The unit suite proves the tool reaches the right verdict
+from a given response; it cannot prove it confirms Webmin.
+
+### Added
+
+- **`--detect-json PATH`** writes a detection run as JSON: the run's overall
+  verdict, per-verdict counts, and every probe with its payload, method, context
+  and evidence. Text output is unchanged. This is the supported way to consume a
+  run programmatically — scraping stdout cannot be made reliable, because a
+  probe payload may contain a literal newline (the newline separator is a real
+  one, so line-oriented parsing splits a payload in half) and the detection path
+  exits 0 whether it confirmed or came back clean.
+- **`tests/bench/` — the coverage benchmark harness.** Each case brings a real
+  vulnerable build up, runs RCEKit as an operator would, checks the verdict, and
+  tears it down; `--markdown` emits the coverage table. Not part of
+  `python -m unittest discover -s tests` — cases need Docker and pull real
+  images — so it runs by hand or in a dedicated job, and exits non-zero if any
+  case fails. Two cases ship, transcribed from `docs/verify-it-yourself.md`:
+  Webmin CVE-2019-15107 and Struts2 S2-001.
+- **A negative control is a required key.** A benchmark without controls measures
+  nothing: a tool that shouted `confirmed` at every target would score full marks
+  on the vulnerable half. Three kinds are supported — a patched build, the same
+  target probed for the wrong class, and a weaker method that must stay below
+  `confirmed` on a target where it happens to be right. The runner refuses four
+  shapes of non-control: no control at all; one expecting `confirmed`; one that
+  runs the identical invocation against an identical target (judged on what it
+  would actually run, so an explicit copy of the vulnerable invocation is caught
+  as well as an omitted one); and one expecting `error` or `nothing-tested`,
+  since both mean the target was never exercised and such a control would stay
+  green with the detection engine entirely broken. Validation and execution
+  share one `control_plan` so they cannot drift.
+- **`overall_detection_verdict`** collapses a run to one verdict, ordered by what
+  an operator must not miss rather than by frequency: one `confirmed` among a
+  hundred negatives is the finding. `error` is reported only when *nothing*
+  reached the target, and a run that built no probes is `nothing-tested` —
+  never `negative`, which would read as "not vulnerable".
+
+### Changed
+
+- `CONTRIBUTING.md` asks for a bench case alongside new detection coverage, and
+  for the README table to state the tier the case actually reached.
+
+### Notes
+
+- The two shipped cases have **not yet been executed through the harness** — it
+  was written where no Docker daemon was available. Their invocations come from
+  a documented, reproduced guide, but the case files themselves are unvalidated;
+  `tests/bench/README.md` says so and flags the one field that is a guess. No
+  README claim was changed to assert benchmark results.
+
+## [2.24.0] — 2026-08-16
+
+The computed value is no longer looked for in the response body alone. A sink
+whose output surfaces anywhere else in the response was reported `negative` — a
+false negative on a class RCEKit already claims to cover, which is worse than a
+missing class. The oracle, the random operands and the control differential are
+unchanged; only the set of places searched is wider.
+
+### Added
+
+- **Whole-response evidence search.** Every confirmation now sweeps the response
+  body, the application response headers, individual cookie values, the redirect
+  target RCEKit actually landed on, the HTTP reason phrase, and each leaf of a
+  parsed JSON body. Real sinks put command output in a debug header or a
+  `Set-Cookie`, and API targets surface an evaluator's result inside a nested
+  error envelope — `{"error": {"detail": "cannot render 2058898001"}}` — where a
+  substring search of the serialised body misses a value the encoder escaped.
+- **The evidence line names the channel that carried the value**, e.g.
+  `target computed 'RK…' in header X-Cmd-Out (random operands, absent from
+  control)`, so the finding stays reproducible by hand. A body-carried
+  confirmation reads exactly as it did before.
+
+### Changed
+
+- **The control differential now covers every channel, not just the body.** A
+  value present anywhere in the payload-free control is not attributable to
+  execution, so it yields `inconclusive` wherever it turned up. This is stricter
+  than comparing only the channel that matched, and it is what keeps a wider
+  search from becoming a looser verdict.
+- **The `file` method's control check covers every channel too**, on the same
+  reasoning: its token is random, so its presence in any control channel means
+  it did not get there by being written and served.
+
+### Security
+
+- **A deeply nested JSON response can no longer silence detection.** Channels are
+  built inside the delivery `try`/`except`, so a `RecursionError` while parsing
+  or walking the body escaped as a network failure: a response that arrived
+  perfectly well was reported "request never reached the target". Measured:
+  every one of the 46 probes in a default `reflected` run turned into `error`,
+  which a target could induce deliberately to hide a live sink behind a thousand
+  nested arrays. Version-independent, though the source moves — CPython 3.12
+  raised the C recursion limit its JSON scanner runs under, so on 3.12/3.13 the
+  parser survives a depth that breaks it on 3.8–3.11 and the recursive leaf walk
+  hit the ordinary Python limit instead. The walk is now iterative and
+  depth-capped, `RecursionError` from the parser costs the JSON channels only,
+  and building channels can never turn a delivered response into a delivery
+  failure.
+- **Transport headers are excluded from the sweep.** `Content-Length`, `Date`,
+  `Age`, `ETag` and their neighbours are generated below the application and can
+  never carry a computed value, but they *are* numeric — and the `expr` probe's
+  expected value is a bare boundary-fenced number. Searching them would let a
+  byte count collide with an arithmetic result and read as execution. Locked in
+  by a test that puts the expected value in `Content-Length` and requires
+  `negative`.
+
+## [2.23.3] — 2026-08-04
+
+Four items from the same review: requests and seconds spent on work that could
+not produce a result. No verdict changes — the lab still confirms 15 of 15
+vulnerable sinks with nothing on the clean five — the run just stops paying for
+probes that were structurally unable to confirm.
+
+### Changed
+
+- **The `awk` probe is no longer sent into a context that wraps the payload in
+  quotes.** It carries double quotes, so in `attribute` the quote closed early
+  and the rest was not a command: 5 requests per carrier that could only ever
+  come back negative. Measured on a verbose shell sink, that shape confirmed 8
+  times in `raw` and 0 times in `attribute`. Break-out contexts such as
+  `shell_double_quoted` *close* the sink's quote and comment its tail, so they
+  still get it. The same guard covers the PowerShell out-of-band shape.
+- **The timing screen runs in two waves.** Every delayed screen probe costs a
+  real sleep, so screening all five separators up front spent `5 × base` seconds
+  on every carrier, including the ones that cannot break out at all. `; ` and
+  `| ` are screened first and the rest only if neither delayed — a sink that
+  filters both is still swept, it is just no longer the price everyone pays.
+- **The out-of-band callback window is no longer paid per carrier.** Callbacks
+  land in a burst once the channel works, so a target that has not produced one
+  across every probe fired so far is not going to. The first carrier still gets
+  the full window, so a target that does call back is never cut short before its
+  first hit. On a clean target with the default carriers this was 30s of pure
+  waiting; it is now ~12s.
+- **`--probe-depth` documents what it does on Windows**, which is nothing:
+  `cmd.exe` has no `#` comment, no `${IFS}` and no `awk`, so both depths send
+  the single `set /a` probe. The docs promised three extra shapes per sink
+  without that caveat.
+
+## [2.23.2] — 2026-08-04
+
+Three findings from a review of the detection work in 2.22.0 and 2.23.0. All
+three are the same shape: the run said something that was not true — about what
+it had done, about what it had looked for, or about which channel was live.
+
+### Fixed
+
+- **`--methods oob` ignored `--verify-active-risk`.** Detection methods build
+  their own probes and so bypass every corpus-level safety filter. That was
+  harmless while every method was inert, but this one makes the target open
+  outbound connections — and the same run printed *"low-impact (safe) payloads
+  only; pass `--verify-active-risk intrusive` to also fire … OOB"* and then fired
+  OOB anyway. It now needs `--verify-active-risk intrusive`, the same tier that
+  holds back the corpus OOB payloads, and refuses before the listener binds.
+- **`--probe-depth quick` silently narrowed the timing separator screen to
+  `; `.** That put back the exact blind spot the screen was added to remove, so
+  a sink that merely filters `;` reported negative — and only for the operator
+  who chose `quick` to be gentle on a rate-limited target. Both depths now screen
+  every candidate separator; `--probe-depth` governs probe *shapes* only, and
+  `--separators` remains the way to narrow break-outs deliberately.
+- **The DNS out-of-band probes could not call back on the default port, and
+  nothing said so.** A DNS callback travels the real resolver hierarchy, so it
+  only arrives if the listener *is* the authority for the OOB domain — port 53
+  plus NS delegation. On `--listen-dns-port 5335` the DNS shapes were still sent,
+  never fired, and the startup line reported `DNS :5335` with no caveat. Since
+  most of the shapes are DNS ones — a resolver is often the only egress a
+  hardened target has — the silence was expensive. RCEKit now says which channel
+  is live.
+- The blind-sink advice added in 2.23.0 suggested an `oob` command without the
+  risk flag, which the gate above would refuse. Naming a command the tool then
+  declines to run is a small version of the same problem, so it now spells out
+  `--verify-active-risk intrusive`.
+
+## [2.23.1] — 2026-08-03
+
+### Added
+
+- **[Verify it yourself](docs/verify-it-yourself.md)** — reproduce the README's
+  confirmations locally against dockerised [vulhub](https://github.com/vulhub/vulhub)
+  targets. Webmin CVE-2019-15107 driven from a captured request (`reflected` →
+  `confirmed`, then `time` → `needs-review` on the *same* sink, which is the
+  clearest demonstration that the tiers are not merged), and Struts2 S2-001
+  (`eval` confirms, `reflected` does not, on a target where both were tried).
+
+  Log4Shell is documented as an advanced case rather than a five-minute one: its
+  sink is a JNDI lookup inside a logging library, so `--methods oob` does not
+  apply — that method builds shell probes for shell-capable environments. The
+  `${jndi:…}` payloads come from the `oob` *category* with the listener
+  correlating the callback, and the token rides in a DNS label, which needs a
+  delegated domain. Saying so is cheaper than a reader discovering it mid-demo.
+
+## [2.23.0] — 2026-08-03
+
+The three sinks v2.22.0 still could not reach. One was a real gap in the probe
+set; the other two were a reporting problem, not a detection one. With both
+closed, a single `--methods reflected,eval,oob` run confirms **all fifteen**
+vulnerable sinks in the lab and still reports nothing on any of the five clean
+ones.
+
+### Added
+
+- **A space-free probe, sent at both probe depths.** Stripping spaces is a filter
+  of the same family as stripping `;` — it looks like it disarms command
+  injection and does not, because `${IFS}` is a space as far as the shell is
+  concerned. Every other probe carries a space, so that one filter silenced all
+  of them and the sink was only reachable if the operator thought to pass
+  `--evade low`. The separator's trailing space is trimmed with it (`;echo…`, not
+  `; echo…`); the newline separator is unaffected. It costs one shape, so it is
+  not part of the `--probe-depth` trade-off, and it is skipped under
+  `--evade low`, which already applies the same transform everywhere.
+- **Guidance when every in-band probe comes back negative.** A results-based
+  method cannot confirm a sink that returns no output — there is nowhere for the
+  computed value to appear — so that negative is not evidence the target is
+  clean. A run of `reflected`/`eval` alone that confirms nothing now says exactly
+  that and names the methods that could still reach a blind sink, with the flags
+  each one needs. It is suppressed once a blind-capable method has already run,
+  and the `file` line is dropped once a web root is known.
+
+## [2.22.0] — 2026-08-03
+
+Detection coverage. Measured against a lab of twenty sinks — fifteen genuinely
+vulnerable, five deliberately clean — the results-based methods went from
+confirming 8 of the 15 to confirming 12, with no new false positives on any of
+the clean ones.
+
+### ⚠️ A blind-timing candidate could be pure latency drift
+
+`--methods time` fired its probes in a fixed ascending delay order
+(`0,0,N,N,2N,2N`), which makes the injected delay collinear with the request
+index. A target that simply gets **slower during the run** — progressive load, a
+rate limiter backing off, a filling log — therefore produced a textbook-perfect
+linear fit while being entirely un-injectable. In the lab this reproduced on 8
+of 8 runs against a sink with no command execution anywhere in it.
+
+The probe order is now randomised, and the request index enters the regression
+as a nuisance term, so drift loads onto a drift coefficient instead of
+masquerading as a sleep. The same lab sink now reports negative on 9 of 9 runs,
+with every genuine timing detection preserved. If you have a `needs-review`
+timing candidate from an earlier version against a target that was under load,
+it is worth re-running.
+
+### Added
+
+- **`--methods oob`** — out-of-band detection, the first `confirmed`-tier method
+  for a sink that returns nothing *and* has no writable web root. Starts the
+  built-in HTTP+DNS listener in-process and asks the target to resolve or fetch
+  `<token>.<oob-host>`; a callback carrying a token the target could only have
+  learned by running the command is proof of execution. Each probe gets its own
+  token, so the finding names the break-out that actually worked. One shape puts
+  a computed value in the DNS label, so the callback proves the shell evaluated
+  arithmetic rather than merely resolving a name. Requires `--oob-host`, since
+  it makes the target open outbound connections.
+- **`--probe-depth quick|full`** (default `full`) — trades requests for
+  coverage. `full` adds three probe shapes, each aimed at a filter that silenced
+  the canonical ones: substitution-free (`awk`, bare `expr`) for sinks that strip
+  `$(` and backticks; keyword-diverse (`awk`) for filters on `echo`/`expr`; and
+  comment-terminated (`… #`) for applications that append a redirect, extra
+  arguments or a pipe after the injection point. `quick` keeps the old probe set
+  at roughly half the requests.
+
+### Fixed
+
+- **A `ping '<input>'` sink could not be detected at all.** The
+  `shell_single_quoted`/`shell_double_quoted` contexts exist precisely for input
+  interpolated inside quotes, but they are not in `default_contexts`, so no
+  record carried them and the detection engine never tried them — the one sink
+  shape they exist for was the one shape that always reported clean. They are now
+  probed by default, and skipped when `--contexts` names a selection explicitly.
+- **`--methods time` reported a `;`-filtering sink as negative.** A regression
+  blends its probes into one measurement, so it could not sweep separators the
+  way the results-based methods do and was locked to `; ` alone — while
+  `| sleep 3` delayed on the same sink. It now screens every candidate separator
+  with one cheap probe each, then runs the regression through whichever one
+  actually delayed.
+- **A trailing redirect or pipe in the sink hid a working probe.**
+  `<cmd> <input> 2>/dev/null` and `<cmd> <input> | grep …` swallow the probe's
+  output, so it executed and still read as negative. The comment-terminated
+  shapes comment that tail out.
+
+## [2.21.1] — 2026-08-02
+
+First release since v2.15.2. The headline is not a new feature — it is that
+detection is now correct in cases where it previously was not.
+
+### ⚠️ Re-check findings from v2.15.2 and earlier
+
+**A reflection could be reported as `confirmed`.** The paired same-token control
+in `run_verification` was gated on a plain `re.search`, while the verdict itself
+used the encoding-aware search. A target that only echoes input but wraps its
+output — base64, hex, URL- or HTML-encoded — skipped the control entirely and was
+reported as proven execution: precisely the case the encoding-aware search was
+added for. If you ran an earlier version against a target that encodes its
+responses, a `confirmed` verdict from that run is worth re-testing.
+
+### Fixed — false negatives on exploitable targets
+
+- **Separator sweep.** Shell probes always broke out with a single hardcoded
+  `; `, so a sink that strips `;` — the most common partial mitigation there is,
+  and one that stops nothing on its own — defeated every probe. Measured against
+  nine deliberately vulnerable local sinks, detection was correct on 5 of 9;
+  three of the four misses were exploitable targets reported clean. Probes now
+  sweep `; `, `| `, `|| `, `&& ` and a newline, narrowable with `--separators`.
+- **Language runtimes.** An environment names what runs the application, not what
+  executes the injected command: PHP's `system()`, Python's `os.system()`,
+  Node's `child_process.exec()`, Ruby's `system()`, Perl's backticks and Go's
+  `os/exec` all hand the string to `/bin/sh`. Scoping a run to the language the
+  application is written in — the natural thing to do — used to send no shell
+  probes at all.
+- **Whole-command sinks.** `--sink-raw` sends probes as bare commands for sinks
+  that execute the input as the entire command (`qx/$input/`, `sh -c "$input"`),
+  where a leading `;` is a syntax error that guaranteed a false negative.
+- **Captured requests.** A trailing newline in a saved request body is no longer
+  sent as part of the body.
+
+### Fixed — a failed request is not a clean result
+
+- A request that never reached the target is reported `error`, not `negative`.
+- Runs that build no probes at all exit non-zero and say so, instead of ending
+  in silence and exit 0 — which read exactly like a target that came back clean.
+- The OOB DNS listener no longer dies on a malformed query, and write failures
+  surface instead of being swallowed.
+
+### Fixed — safety and audit
+
+- **Multi-step chains now carry the same safeguards as single requests.** The
+  chain path delivered to live targets without sink-shape filters, destructive
+  hold-back or a pre-flight plan, so `--verify-active-risk stateful` fired
+  persistence and irreversible file operations that `--verify-url` refuses to
+  send without `--verify-allow-destructive`. Both paths now share one hold-back
+  and print the same plan.
+- **The audit trail redacts credentials** — it records that a credential header
+  was sent, never its value.
+- A capture carrying `Authorization` or `Cookie` over plain `http` is flagged
+  before anything is sent.
+
+### Added
+
+- **The payload corpus is embedded in `rcekit.py`**, so the single file runs on
+  its own — a jump box, an air-gapped host, a bare `curl` of the raw script.
+  Resolution order is `--template-file` → `templates/payloads.json` beside the
+  script → the built-in copy, and falling back to the built-in copy is
+  announced. A corpus that exists but does not parse still hard-fails: that
+  check exists for truncated and tampered corpora. `tools/embed_corpus.py`
+  regenerates the embedded copy, and the test suite fails if the two drift.
+- **`--insecure`** skips TLS verification for internal targets with self-signed
+  or mismatched certificates — opt-in and explicit, like `curl -k`.
+- **`--sink-raw`** for whole-command injection sinks, also readable from a
+  target profile.
+- **`--separators`** to narrow the break-out sweep once the sink's shape is
+  known.
+- **Documentation split into a task-oriented tree.** The README is half its
+  former length and now leads with what RCEKit is for:
+  [field guide](docs/guide.md) (worked examples by situation),
+  [payload generation & exports](docs/generation.md), and
+  [reference](docs/reference.md) (every flag grouped by task, plus the full
+  taxonomies and exit codes).
+- **A "How RCEKit compares" section** covering commix, SSTImap, Nuclei and
+  interactsh, with every claim traceable to that project's own documentation.
+- Four confirmation demos against real, publicly documented CVEs (Webmin
+  CVE-2019-15107, Struts2 S2-001, Log4Shell CVE-2021-44228).
+
+### Changed
+
+- **Expect more requests per run.** The separator sweep and the language-runtime
+  fix both widen the probe set. Narrow with `--separators`, `--contexts` and
+  `--environments` once the sink's shape is known.
+- **`--doctor` output.** Its first line now names the corpus in use
+  (`corpus: …`) rather than a path (`template: …`), since the corpus is no
+  longer necessarily a file, and `[ok] file loaded and parsed` is now
+  `[ok] corpus loaded and parsed`.
+
+No breaking changes to the CLI, output formats, or the template schema.
+Standard library only, Python 3.8–3.13.
+
+## Earlier releases
+
+Release notes for these live on the
+[Releases page](https://github.com/kabiri-labs/rcekit/releases); they predate
+this file and have not been restated here.
+
+- **[2.15.2]** — Multi-method RCE detection &amp; confirmation
+- **[2.7.0]**
+- **[2.1.0]**
+
+
+
+[Unreleased]: https://github.com/kabiri-labs/rcekit/compare/v2.36.0...HEAD
+[3.0.0]: https://github.com/kabiri-labs/rcekit/compare/v2.45.7...v3.0.0
+[2.45.7]: https://github.com/kabiri-labs/rcekit/compare/v2.45.6...v2.45.7
+[2.45.6]: https://github.com/kabiri-labs/rcekit/compare/v2.45.5...v2.45.6
+[2.45.5]: https://github.com/kabiri-labs/rcekit/compare/v2.45.4...v2.45.5
+[2.45.4]: https://github.com/kabiri-labs/rcekit/compare/v2.45.3...v2.45.4
+[2.45.3]: https://github.com/kabiri-labs/rcekit/compare/v2.45.2...v2.45.3
+[2.45.2]: https://github.com/kabiri-labs/rcekit/compare/v2.45.1...v2.45.2
+[2.45.1]: https://github.com/kabiri-labs/rcekit/compare/v2.45.0...v2.45.1
+[2.45.0]: https://github.com/kabiri-labs/rcekit/compare/v2.44.0...v2.45.0
+[2.44.0]: https://github.com/kabiri-labs/rcekit/compare/v2.43.0...v2.44.0
+[2.43.0]: https://github.com/kabiri-labs/rcekit/compare/v2.42.0...v2.43.0
+[2.42.0]: https://github.com/kabiri-labs/rcekit/compare/v2.41.0...v2.42.0
+[2.41.0]: https://github.com/kabiri-labs/rcekit/compare/v2.40.0...v2.41.0
+[2.40.0]: https://github.com/kabiri-labs/rcekit/compare/v2.39.0...v2.40.0
+[2.39.0]: https://github.com/kabiri-labs/rcekit/compare/v2.38.0...v2.39.0
+[2.38.0]: https://github.com/kabiri-labs/rcekit/compare/v2.37.0...v2.38.0
+[2.37.0]: https://github.com/kabiri-labs/rcekit/compare/v2.36.0...v2.37.0
+[2.36.0]: https://github.com/kabiri-labs/rcekit/compare/v2.35.5...v2.36.0
+[2.35.5]: https://github.com/kabiri-labs/rcekit/compare/v2.35.4...v2.35.5
+[2.35.4]: https://github.com/kabiri-labs/rcekit/compare/v2.35.3...v2.35.4
+[2.35.3]: https://github.com/kabiri-labs/rcekit/compare/v2.35.2...v2.35.3
+[2.35.2]: https://github.com/kabiri-labs/rcekit/compare/v2.35.1...v2.35.2
+[2.35.1]: https://github.com/kabiri-labs/rcekit/compare/v2.35.0...v2.35.1
+[2.35.0]: https://github.com/kabiri-labs/rcekit/compare/v2.34.1...v2.35.0
+[2.34.1]: https://github.com/kabiri-labs/rcekit/compare/v2.34.0...v2.34.1
+[2.34.0]: https://github.com/kabiri-labs/rcekit/compare/v2.33.0...v2.34.0
+[2.33.0]: https://github.com/kabiri-labs/rcekit/compare/v2.32.0...v2.33.0
+[2.32.0]: https://github.com/kabiri-labs/rcekit/compare/v2.31.0...v2.32.0
+[2.31.0]: https://github.com/kabiri-labs/rcekit/compare/v2.30.0...v2.31.0
+[2.30.0]: https://github.com/kabiri-labs/rcekit/compare/v2.29.0...v2.30.0
+[2.29.0]: https://github.com/kabiri-labs/rcekit/compare/v2.28.0...v2.29.0
+[2.28.0]: https://github.com/kabiri-labs/rcekit/compare/v2.27.0...v2.28.0
+[2.27.0]: https://github.com/kabiri-labs/rcekit/compare/v2.26.0...v2.27.0
+[2.26.0]: https://github.com/kabiri-labs/rcekit/compare/v2.25.0...v2.26.0
+[2.25.0]: https://github.com/kabiri-labs/rcekit/compare/v2.24.0...v2.25.0
+[2.24.0]: https://github.com/kabiri-labs/rcekit/compare/v2.23.3...v2.24.0
+[2.23.3]: https://github.com/kabiri-labs/rcekit/compare/v2.23.2...v2.23.3
+[2.23.2]: https://github.com/kabiri-labs/rcekit/compare/v2.23.1...v2.23.2
+[2.23.1]: https://github.com/kabiri-labs/rcekit/compare/v2.23.0...v2.23.1
+[2.23.0]: https://github.com/kabiri-labs/rcekit/compare/v2.22.0...v2.23.0
+[2.22.0]: https://github.com/kabiri-labs/rcekit/compare/v2.21.1...v2.22.0
+[2.21.1]: https://github.com/kabiri-labs/rcekit/compare/v2.15.2...v2.21.1
+[2.15.2]: https://github.com/kabiri-labs/rcekit/releases/tag/v2.15.2
+[2.7.0]: https://github.com/kabiri-labs/rcekit/releases/tag/v2.7.0
+[2.1.0]: https://github.com/kabiri-labs/rcekit/releases/tag/v2.1.0