sourcecode 3.2.0__tar.gz → 3.2.2__tar.gz

{sourcecode-3.2.0 → sourcecode-3.2.2}/CHANGELOG.md

@@ -1,5 +1,305 @@
 # Changelog
 
+## [Unreleased]
+
+### Changed
+
+- **Telemetry is opt-in** (P-1). It was opt-out; with no explicit choice on record, nothing is
+  now collected or transmitted — not on the first run, not in CI, nowhere. **An explicit choice
+  you made before is honoured unchanged**; if you never chose, telemetry is off.
+  - The default was not a preference. Field evaluation #3 was the first audit of regulated
+    third-party code (public-sector health data), where the reviewer must be able to state that
+    nothing left the machine — and *"disable it after the first run"* cannot give that
+    guarantee. It was recorded as a **procurement** blocker, not an analysis defect.
+  - The CI special case is gone along with the default it existed to override. CI is still
+    detected for exactly one purpose: skipping a first-run notice that has no terminal to
+    appear on.
+  - The first-run notice is now an **invitation**: it says telemetry is off, states what *would*
+    be collected if you enabled it, and names `ask telemetry enable`. It changes no setting.
+  - `ask telemetry status` distinguishes **"no choice recorded"** from **"you chose disabled"**
+    (`stored_choice()` returns `None`, never a silent `False`) — a buyer auditing this needs to
+    know which of the two they have.
+
+## [3.2.2] — 2026-07-29
+
+**M7.5 complete — the product now says what it promises.** Three things a user could only
+infer are written down and enforced: which of their commands a warm cache actually helps, what
+a cold run costs where they meet it, and **what each command's output is worth relying on**.
+Documenting the first of those found the cache had two freshness authorities and was answering
+for trees it had not read — the pattern of this cycle, where writing down the promise is what
+exposes where we do not keep it.
+
+### Added
+
+- **Command tiers, named and published** (C4-6). Every command now carries one of four tiers —
+  `core`, `supported`, `experimental`, `parked` — from a single authority (`cli.COMMAND_TIERS`),
+  rendered by `ask --help`, the README command table and the user guide.
+  - A tier is a **stability promise**: `core` is safe to gate CI on, `experimental` may change
+    shape in a minor, `parked` still works but is no longer developed (empty today, and the help
+    says so rather than leaving the row blank).
+  - It is **not a value ranking** — `posture` is experimental *and* the first command the help
+    recommends — and it is **not the pricing tier**, which gates repository size and never
+    capability. Both distinctions are printed, not implied.
+  - The battery fails if a registered command sits in no tier, if a tier names a command that
+    does not exist, if a command calls itself `[EXPERIMENTAL]` while published in another tier,
+    or if either document drifts from the table.
+  - Which commands *move* between tiers, and whether `parked` disappears from the default help,
+    is the tiering work in M9. This release names the vocabulary before it is used against a
+    command.
+- **The cache model is published, per command, and measured.** `ask cache model` (`--json`,
+  `--markdown`) and [docs/CACHE.md](docs/CACHE.md) answer the question field evaluation #3 could
+  not: *which of my commands does a warm actually help?* One rule for invalidation — every layer
+  keys on the exact tree state — then, per command, whether a warm gives it **the answer**, only
+  **the shared work**, or **nothing**, with the layers it reads and a measured figure.
+  - Which layers each command reads was derived by instrumenting the layers, not by reading the
+    code and hoping. What a warm is worth was then timed **per command in isolation**: cleared
+    caches, `cache warm`, then the command. Timed in a batch instead, one command warms the next
+    and the table lies — `impact-chain` looked unaffected by a warm in a batch and goes 9.9 s →
+    1.7 s in isolation.
+  - Two measured results contradict the obvious assumption and are published as such: a warm
+    buys **`endpoints` nothing** (2.8 s → 2.9 s) and **`migrate-check` nothing** (4.8 s → 4.8 s),
+    and on a small diff it makes `review-pr` *slower* (1.1 s → 2.3 s) because loading the warmed
+    IR costs more than the work it saves.
+  - The battery fails if a registered command has no row, if the model names a command that does
+    not exist, if a row claims a warm helps while reading no layer a warm builds, or if
+    `docs/CACHE.md` has drifted from the registry. This is the `--help` lesson (C4-1) applied: a
+    curated list beside a generated one drifts.
+- **`cache warm` states its cost and its limits when it finishes** (C4-5): elapsed time, what it
+  warmed, what it did **not** — `prepare-context` task answers, runs whose analysis flags differ,
+  and the two commands measured to gain nothing — plus the line that matters in CI: without a
+  persisted `~/.sourcecode`, every pipeline pays that time again. `cache warm --help` says it
+  before the run, not after.
+
+### Fixed
+
+- **One authority for cache freshness: the tree state, never the committed HEAD.** The
+  snapshot cache behind the root command keyed on the HEAD sha alone, so it answered for a
+  tree it had not read. Reproduced on spring-petclinic: add `spring-boot-starter-data-ldap`
+  to a tracked `pom.xml`, do not commit, run `ask --compact` again — the pre-edit analysis
+  comes back from `L2_view` with `is_stale: false`, the new dependency invisible. The context
+  cache sitting next to it keyed on the exact tree and invalidated correctly, which is why
+  field evaluation #3 could only describe the behaviour as *"el modelo de invalidación no
+  está documentado y no es predecible"*: there were two models.
+  - `cache.worktree_signature` is now the single authority every layer keys on — the root
+    command's L1/L2 snapshots, the `prepare-context` task cache, the shared Canonical IR, and
+    the RIS bootstrap. One rule describes all of them: **any change to the analysed files
+    invalidates, committed or not.** A clean tree keys on the sha, so the ordinary repeat run
+    still hits (BroadleafCommerce: 18.1 s cold → 0.3 s warm, unchanged; the signature itself
+    costs ~50 ms on that tree, ~140 ms on keycloak).
+  - Untracked files are covered by size and mtime: they appear in no diff, so a new file — or
+    an edit to one — was invisible to a signature built from `git diff` alone.
+  - A tree git cannot describe (no repository, no commit) gets a fingerprint over the size and
+    mtime of every non-ignored file. Before, such trees used a **constant** key, so the cache
+    answered forever with the first analysis it ever ran; the context cache disabled itself
+    instead. Both now cache and both now invalidate.
+  - `--scope`-style subdirectory runs sign only their own subtree: an edit in a sibling module
+    no longer invalidates the module actually analysed.
+- **RIS staleness answers the question it is asked.** `get_cold_start_context` reported
+  `stale: false` for any tree sharing the commit the snapshot was taken at, while carrying
+  `has_uncommitted_changes: true` in the same object — two facts about one thing, disagreeing.
+  The snapshot now records the tree it describes (`metadata.tree_signature`), and the
+  `prepare-context` RIS fast path serves `onboard`/`explain` only when that tree is still the
+  one on disk. Snapshots written before this release fall back to the old pair of facts.
+- **`prepare-context --all` and `--include-config` are in the cache key.** They changed the
+  answer and were not keyed; the collision could not surface only because the key demanded a
+  git sha, and non-git trees cached nothing at all. Now that every tree has a signature, an
+  unkeyed flag would be served the other flag's answer.
+
+## [3.2.1] — 2026-07-29
+
+**M7 complete — the last four seams, the fact registry, and a `--help` that leads with the
+product.** Every fact ADR-0008 named now has one authority, a registered consumer list and an
+assertion that keeps its emitters agreeing. Two of the four seams closed here were found by
+field evaluation #3 *between the two commands it valued most*, and reproducing them showed
+both were worse than reported.
+
+**Upgrade note.** Three published numbers change value. None is a schema break; all are
+corrections, and a pinned threshold may move:
+
+- `modernize.summary.statically_unreferenced` grows sharply — the population it is measured
+  over never matched its own definition (keycloak 14 → 108, BroadleafCommerce 10 → 36,
+  alfresco 13 → 88). `summary.high_coupling_nodes` likewise stops reporting the display cap
+  of 20 and reports the measured count.
+- `endpoints --limit N` no longer reports `total: N`. `total` and the security counters
+  describe the endpoints the *filters* selected; the new `shown` says how many are listed.
+- `posture` verdicts change on repositories that declare `spring.mvc.servlet.path`: rules are
+  matched against the URL the filter chain actually receives, so endpoints previously reported
+  `no_rule_matched` — or, with a catch-all behind the rule, `permit_all` — now carry the
+  decision the matcher makes. Repositories without a declared servlet path are byte-identical.
+
+New keys (`summary.reference_status`, `effective_access.path_basis`, `matched_path`, `shown`,
+and `registries` in both `ask schema` and `ask version`) are additive.
+
+### Changed
+
+- **`--help` no longer introduces the product with its weakest 20 %.** The command table was
+  ordered by the order commands are registered in the source, so a reader met
+  `prepare-context` and `repo-ir` — the two surfaces field evaluation #3 scored lowest —
+  before `posture` and `endpoints`, which the same evaluation called "the product". Commands
+  are now grouped into ordered panels, led by **Java/Spring analysis — start here**
+  (`posture`, `endpoints`, `spring-audit`, `migrate-check`), then *Change and risk*, *Context
+  for AI agents*, *Setup and inspection*, and *Experimental*. Nothing is hidden and nothing is
+  removed; the ordering is one table (`cli.HELP_PANELS`) and the battery fails if a registered
+  command is missing from it, if the table names a command that does not exist, or if the core
+  four stop leading the rendering.
+- **The hand-written "Also available:" catalogue is gone.** A curated list beside a generated
+  one drifts, and it did: it is how `endpoints` shipped while appearing in no `--help` list at
+  all, leaving an evaluator to find the command they valued most inside another command's JSON.
+- **`ask spring-audit --help` now says which gate it is.** `--ci` fails on *any* finding,
+  including debt that was already there — the field read it as baseline-relative and concluded
+  the gate was unusable on a legacy repository. The help now states the gate is absolute and
+  names the baseline-relative one (`ask verify --fail-on new`, `ask pr-impact --fail-on`) and
+  `ask baseline` for tracking the debt. A test asserts those commands and flags exist.
+
+### Added
+
+- **Invariant I-3 in the battery: every registered fact declares what it says when it cannot
+  decide.** ADR-0008 R9 is easy to state and easy to lose — `statically_unreferenced: 0` over
+  3 374 classes is what losing it costs. Each registry entry now carries `unresolved_answer`
+  (the outcome the fact emits when undecidable) plus the reason, or `null` with an explanation
+  of why the fact is total. A declared outcome must exist in the authority that derives it, so
+  the registry cannot document an `unknown` nothing emits — a check that immediately corrected
+  two entries where the unresolved answer belongs to the consumer, not the authority.
+
+- **Invariant I-2 in the battery: a published count may not come from a truncated list.**
+  A static audit over every module (`tests/test_invariant_counts_from_measurement.py`): inside
+  one function, a name bound to a slice and then measured with `len()` in a payload position
+  fails the suite. Display sizes stay legal under an explicit key (`shown`, `omitted`, …),
+  because "how many did you show me" is a different question from "how many are there". The
+  audit is itself guarded against being vacuous.
+  - It found two live ones. `modernize.summary.high_coupling_nodes` reported the display cap
+    (20) as the number of high-coupling classes — a repository with 300 of them said 20; the
+    exact count now ships alongside `high_coupling_nodes_shown`. And `endpoints --limit N`
+    reported `total: N`: `--limit` cuts the rendering, not the population.
+  - **Value change:** with `--limit`, `endpoints` now reports `total` (and the security
+    counters) for the endpoints the *filters* selected, plus a new `shown` for how many are
+    listed. Before, all of them collapsed to the limit. Filters (`--path-prefix`,
+    `--controller`) genuinely change the population and still move `total`.
+
+- **The fact registry ADR-0008 asks for (R11), shipped inside the package and enforced.**
+  `src/sourcecode/facts/registry.json` names every fact that has a single authority: what it
+  means, which `module:symbol` derives it, which modules may emit it, and the parity assertion
+  that keeps those emitters agreeing. Read it from an installed version with
+  `ask schema facts-v1` — a release states its own contract instead of pointing at
+  documentation that can drift from it. No new command: `ask schema` lists it under
+  `registries`, beside the output `schemas`.
+  - The registry is checked, not asserted: the battery fails if an authority does not exist,
+    if a registered consumer does not bind to it (R13 — the LDAP defect survived two releases
+    exactly because a consumer was never rebound), or if a fact carries no parity test (R12).
+    Writing it found two claims of mine that were untrue and one surface — the endpoint
+    security verdict — whose only real consumer is `impact`.
+  - Eleven facts registered, one per M7 seam plus the two each seam split apart.
+
+### Fixed
+
+- **A repository with no tests at all reported `has_tests: true` and "N test files".** Third
+  field evaluation: a Spring monolith whose `src/test` does not exist was described as having
+  4 test files, because production classes living in a package named `test` matched a
+  substring rule. The same document's `analysis_gaps` contradicted it — one document, two
+  derivations of one fact.
+  - New authority `test_sources`: a file is a test source when it sits under a **declared
+    test source root** (`src/test`, `src/it`, `src/androidTest`, `src/testFixtures`, or a
+    repository-level `tests/`, `spec/`, `__tests__/`) or matches an ecosystem **naming
+    convention** (`FooTest.java`, `test_foo.py`, `foo.spec.ts`, `foo_test.go`). A `test`
+    package **inside a main source root is production code** — the build tool's own layout
+    outranks any name.
+  - Five consumers now bind to it instead of re-deriving: `signals.has_tests`, the
+    `analysis_gaps` backend-coverage check (which now also states its basis),
+    `metrics_analyzer.is_test_file`, `prepare-context`'s test/source split, and the review-pr
+    coverage risk.
+  - **Value change:** files under a *main* source root in a directory named `test`/`tests`
+    stop counting as tests. Measured on the fleet: keycloak 2 220 → 1 777 Java test files
+    (443 files under `testsuite/…/src/main/java/…`), BroadleafCommerce 201 → 197,
+    spring-petclinic and openmrs-core unchanged.
+- **`review-pr` reported `test_coverage_risk: "low"` for a diff containing no source files.**
+  In a repository with no tests, a documentation-only PR read as "covered" — the traffic light
+  inverted exactly where a reviewer relies on it. The surface now emits `unknown` with the
+  reason (`no source files changed in this diff`, or `repository declares no test sources`),
+  plus `changed_source_files` and `repository_has_test_sources`. `low` is now reachable only
+  when changed source files were checked and matched (ADR-0008 R9).
+
+- **A class was reported as conditional on a profile *and* on its negation, contradicting
+  `ask posture` in the same run.** Found in the third field evaluation (3.2.0): a
+  configuration class declaring `@Profile("!m3")` — active exactly when `m3` is **absent** —
+  appeared in `signals.spring_profiles.profile_conditional_beans` under `m3` as well. Read as
+  a security fact, it inverts the answer.
+  - Root cause, reproduced: the source scan attributed every `@Profile` to *the next type
+    declaration ahead of it*, falling back to the file name when there was none. A
+  `@Profile("m3")` on a `@Bean` **method** inside that class was therefore attributed to the
+    class itself, next to the class-level `@Profile("!m3")`.
+  - Attribution is now the declaration the annotation actually annotates: only annotations
+    and modifiers may stand between it and a type declaration; a member-level annotation is
+    reported as `Type#member`, and an annotation whose target cannot be read is **not
+    reported at all** rather than attributed to a guess (ADR-0008 R9).
+  - **Value change:** member-level declarations move from the enclosing type to the member.
+    On the reference fleet this affects one symbol (`spring-petclinic`:
+    `MysqlTestApplication` → `MysqlTestApplication#container`); on repositories that put
+    `@Profile` on `@Bean` methods it affects each such method.
+  - New in the payload: `signals.spring_profiles.conditional_beans[]` with `symbol`,
+    `profiles`, `expression` (as written) and `source_file`. The token map
+    (`profile_conditional_beans`) buckets `@Profile("a & !b")` under `a` and `!b` with the
+    operator lost; the list keeps the expression, and both are now projections of **one**
+    derivation, so they cannot disagree.
+  - Parity assertion added: no symbol may be reported as conditional on a profile that
+    `posture` resolves it inactive under (invariant I-1).
+
+- **`ask modernize` reported `statically_unreferenced: 0` across a 3 374-class monolith.**
+  Read as "there is no unreferenced code here", when what it meant was "a check that cannot
+  see framework dispatch found nothing" — the one place the product broke its own rule about
+  never publishing a number where the honest answer is `unknown` (ADR-0008 R9). Two causes:
+  - The population never matched the published definition. The note said "zero static
+    callers", the code examined only types with **no edges in either direction** — so a type
+    that calls something but is called by nothing was never a candidate at all.
+  - The surface had two buckets and the question has three answers.
+  - One authority now derives it, `reference_facts`, as a partition that sums to the classes
+    examined: `referenced` (an incoming edge, or another source naming it) /
+    `unknown_framework_dispatch` (no static caller, but a published annotation, an entry-point
+    signature, or the name wired from a configuration file) / `no_static_callers` (neither —
+    **absence of evidence, never "dead code"**). `summary.reference_status` carries the exact
+    counts and a `statement` that explains a zero instead of publishing it bare.
+  - Counts come from the measurement, not from `len()` of the lists, which are cut to 20 for
+    display (R5). Every listed type carries the `basis` for being there.
+  - Dispatch detection extended with published stereotype, persistence and entry-point
+    vocabulary (`@Component`/`@Service`/`@Repository`/`@Controller`/`@RestController`/
+    `@Configuration`/`@Entity`/`@Embeddable`/`@MappedSuperclass`/`@Aspect`/`@WebServlet`/
+    `@WebFilter`/`@WebListener`/`@SpringBootApplication`, `public static void main`), and the
+    configuration-name scan now matches whole names — `Order` inside `OrderItemImpl` was
+    marking live-looking classes dispatched.
+  - **Value change:** the candidate population is much larger, so the counts move.
+    keycloak 14 → 108 `no_static_callers`, BroadleafCommerce 10 → 36, alfresco 13 → 88;
+    spring-petclinic stays 0 and now says why (its one callerless type is an `@Entity`, so
+    the answer is `unknown`). The scan is also faster than the one it replaces: one repository
+    walk instead of two, and fixed patterns intersected with the candidate set rather than an
+    alternation over it (keycloak 2.5 s → 1.2 s).
+
+- **`ask posture` decided endpoint access against a URL the application never serves.** A
+  repository declaring `spring.mvc.servlet.path=/api` is served at `/api/orders`, and that is
+  the URL its filter chain matches — but the chain rules were compared against the
+  mapping-relative `/orders`. A rule reading `.antMatchers("/api/**").authenticated()`
+  therefore covered nothing. Reported from the field as 2 632 endpoints falsely
+  `no_rule_matched`; reproduced in a fixture, it is worse than that — where a catch-all
+  `anyRequest().permitAll()` follows the rule, those endpoints were reported **`permit_all`**,
+  which states a protected endpoint is open.
+  - One authority for the prefix: `deployment_prefix`, the same module `ask endpoints` reads
+    for `effective_path`. `posture` asks it for the part a matcher actually sees — the
+    **servlet path**, resolved for the profile set you asked about, so a servlet path declared
+    only under `prod` applies exactly when `prod` does. The servlet **context** path is
+    excluded on purpose: the container strips it before the filter chain runs.
+  - Matcher semantics are honoured rather than averaged: `antMatchers`/`regexMatchers` match
+    the servlet-path-prefixed URL, `mvcMatchers` matches relative to the servlet mapping, and
+    `requestMatchers` is either one depending on the Spring Security version and classpath —
+    unreadable from source, so a rule covering the request under either reading is honoured.
+  - New in the payload: `effective_access.path_basis` states the servlet path used and where
+    it came from, and each stated endpoint carries `matched_path` in its evidence when the
+    matched URL differs from the endpoint's mapping-relative path.
+  - **Never a guessed URL** (ADR-0008 R9): where a servlet path is declared but cannot be
+    resolved (a `${PLACEHOLDER}`, or documents that disagree), every verdict that depends on a
+    path pattern is reported `undecided` with `decision_if_unprefixed`, instead of computed
+    against an assumed empty prefix.
+  - **Value change:** none on a repository that declares no servlet path — the two readings
+    coincide and the matching is unchanged. Verified byte-identical on spring-petclinic,
+    keycloak, BroadleafCommerce and spaghetti-api, none of which declare one.
+
 ## [3.2.0] — 2026-07-28
 
 **Answer coherence (M7, first four seams).** One fact, one authority (ADR-0008). Four

{sourcecode-3.2.0 → sourcecode-3.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sourcecode
-Version: 3.2.0
+Version: 3.2.2
 Summary: Persistent structural context and ultra-fast repeated analysis for AI coding agents
 License-File: LICENSE
 Keywords: agents,ai,codebase,context,developer-tools,llm
@@ -73,6 +73,19 @@ Cache keyed on content hashes — invalidated only when source changes. On repea
 
 At 0.3s per call, ASK Engine becomes **constant infrastructure** inside agent loops — call it before every edit, every PR review, every test run, without batching or caching manually.
 
+**What a warm actually covers.** `ask cache warm` runs the compact analysis: it rebuilds the
+shared structural layers (L1/L2 + the Repository Intelligence Snapshot + the shared Canonical
+IR) and the **compact view**. Pass `--agent` to warm the agent view as well. Deeper projections
+are separate keys and are *not* covered by either — `--full`, `--env-map` and a raised `--depth`
+recompute on first use, as do most `prepare-context` tasks. Measured on a 3,342-file Spring
+monolith: `cache warm` 103s → `--compact --git-context` 1s (hit), but `--agent --full --env-map
+--depth 20` still 171s (miss).
+
+**One rule for invalidation, and a table that names every command.** Every layer keys on the
+exact tree state: any change to the analysed files invalidates it, committed or not. Which of
+*your* commands a warm helps — the answer, only the shared work, or nothing — is published per
+command: `ask cache model`, or [docs/CACHE.md](docs/CACHE.md).
+
 ---
 
 ## Install
@@ -95,27 +108,55 @@ ask version                    # ask 2.5.1
 
 ## Quickstart
 
+**Start with these four.** In three independent field evaluations they carried most of the
+measured value, and `posture --diff` is the one no evaluator found an equivalent for —
+commercial or open source.
+
 ```bash
-# High-signal structural summary — warm cache ~0.3s, cold 2–10s
-ask --compact
+# What does each profile set ACTUALLY wire — and what changes between them?
+# Resolves conditional beans and the filter chain, then diffs effective endpoint access.
+ask posture /path/to/repo --diff dev:prod
 
-# Blast radius: what breaks if this class changes?
-ask impact OrderService /path/to/repo
+# Every REST endpoint with its effective path (context-path + servlet path resolved),
+# its inferred security policy and a confidence per endpoint.
+ask endpoints /path/to/repo
+
+# Spring semantic audit: transactional anomalies (private @Transactional = silent
+# CGLIB no-op), security surface, request-body validation.
+ask spring-audit /path/to/repo
 
-# Spring Boot 2→3 migration readiness (bounded decision summary)
+# Spring Boot 2→3 readiness: located blockers, per-dimension score, effort estimate.
 ask migrate-check /path/to/repo --compact
+```
 
-# Spring semantic audit: TX anomalies + security surface
-ask spring-audit /path/to/repo
+Then the everyday loop:
+
+```bash
+# High-signal structural summary — warm cache ~0.3s, cold 2–10s
+ask --compact
+
+# Blast radius: what breaks if this class changes?  (target the INTERFACE, not the Impl)
+ask impact OrderService /path/to/repo
 
 # Onboard to an unfamiliar codebase
 ask onboard /path/to/repo
 
 # PR review: risk, test gaps, changed modules
 ask review-pr /path/to/repo --since main
+
+# CI gate on NEW violations only, instead of on pre-existing debt
+ask verify /path/to/repo --capture-baseline   # accept today's debt, once
+ask verify /path/to/repo                      # then: only new violations block
 ```
 
-Full command reference: **[docs/USER_GUIDE.md](docs/USER_GUIDE.md)**.
+> **Adopting a gate on a real codebase.** A repository that starts declaring contracts already
+> violates them somewhere; `ask verify` is **baseline-relative by default** (`--fail-on new`)
+> so the gate survives contact with reality instead of being switched off on day one.
+> `ask baseline capture|diff|trend` is a different thing: versioned architectural metrics over
+> time, for trend reporting rather than blocking.
+
+Full command reference: **[docs/USER_GUIDE.md](docs/USER_GUIDE.md)** · posture in depth:
+**[docs/posture.md](docs/posture.md)**.
 
 ---
 
@@ -148,17 +189,76 @@ Deterministic Spring semantics: transactional anomalies (e.g. `@Transactional` o
 `ask spring-audit` · `ask validation`
 → [reference](docs/USER_GUIDE.md#core-commands)
 
+### 5b · Runtime Posture *(experimental — and the most differentiated thing here)*
+What a profile set **actually wires**: which conditional beans register, which do not, and which conditions could not be decided at all — then the effective endpoint access that follows from the filter chain. `--diff` answers the question nobody else answers in one command: *what changes between `dev` and `prod`, across every endpoint at once.*
+`ask posture` · `ask posture --diff dev:prod` · `ask posture --property k=v`
+→ [posture.md](docs/posture.md)
+
+Unresolved is a first-class outcome: a condition the resolver cannot decide is reported as a
+hole with the condition named, never folded into active or inactive. **A posture answer that
+guesses is a confident security falsehood — the worst failure mode this tool has.**
+
 ### 6 · Developer Workflows
 The everyday loop: diff-based PR review, symptom-driven bug triage, and delta context for continuous agent runs.
 `ask review-pr` · `ask fix-bug` · `ask prepare-context`
 → [reference](docs/USER_GUIDE.md#typical-workflows)
 
 ### 7 · Utilities
-`ask rename-class` (word-boundary Java rename) · `ask chunk-file` (split large files for agents) · `ask cache` (status / warm / clear / freshness)
+`ask rename-class` (word-boundary Java rename) · `ask chunk-file` (split large files for agents) · `ask cache` (status / warm / model / clear / freshness)
 → [reference](docs/USER_GUIDE.md)
 
 ---
 
+## Command tiers
+
+A tier says **what an output is worth relying on** — it is a stability promise, not a value
+ranking, and not the [pricing tier](docs/PRODUCT_TIERS.md) (Free/Pro gates repository size,
+never capability). `posture` is the most differentiated command in the product *and* it is
+experimental: both are true, and they are two different facts.
+
+| Tier | Promise | Commands |
+|---|---|---|
+| **core** | contract stable within a major — safe to gate CI on | `endpoints` · `spring-audit` · `migrate-check` · `impact` · `impact-chain` · `pr-impact` · `verify` |
+| **supported** | maintained; fields are added, never removed without a major | every command not named in another row |
+| **experimental** | shape may change in a minor — do not gate CI on it | `posture` · `archetype` · `retrieve` |
+| **parked** | kept working, no longer developed | none today |
+
+The same table is printed by `ask --help`, and both are generated from one authority
+(`cli.COMMAND_TIERS`) — the battery fails if a command is in no tier, or if this file drifts
+from it.
+
+---
+
+## Every command, in one table
+
+*`ask --help` shows a short header; this is the full surface. If you only read one row, read
+`posture`.*
+
+| Command | Tier | Answers | Note |
+|---|---|---|---|
+| `posture` | experimental | which beans a profile set wires, and how effective endpoint access differs between two sets | the most differentiated capability here |
+| `endpoints` | core | every REST endpoint, effective path, security policy, confidence | Spring MVC + JAX-RS (~65 % recall on JAX-RS sub-resource locators) |
+| `spring-audit` | core | transactional anomalies + security surface + validation gaps | `--ci`, `-f github-comment` |
+| `migrate-check` | core | Boot 2→3 readiness: located blockers, per-dimension score, effort | `--blast-radius` orders the re-test plan |
+| `impact` / `impact-chain` | core | blast radius of a change, to the endpoints it reaches | target the **interface**, not the `Impl` |
+| `pr-impact` | core | the same, scoped to a PR diff | gating command: `--fail-on`, exit codes |
+| `verify` | core | does the repo satisfy its declared contracts, **relative to a baseline** | `.ask/contracts.yml`; exit 0/1/2 |
+| `verify-edit` | supported | did the working-tree edits change runtime behaviour | semantic diff gate for the edit loop |
+| `--compact` / `--agent` | — | bounded structural context for an agent | flags of the root command, not commands: not tiered |
+| `onboard` / `explain` / `cold-start` | supported | orientation in an unfamiliar repo; per-class summary; bootstrap snapshot | |
+| `export` / `repo-ir` / `schema` | supported | tool-agnostic views (C4, module graph, integrations); symbol-level IR; published JSON Schemas | |
+| `modernize` | supported | coupling hubs, cycles, dead zones, refactor candidates | |
+| `review-pr` / `fix-bug` / `prepare-context` | supported | diff review, symptom triage, task-shaped context | |
+| `plan` / `compare` / `delta` / `contract-diff` | supported | what to review for a change; candidates by measured cost; outcome of a change; public-contract break | no verdicts, measured cost only |
+| `validation` | supported | request-body validation coverage and gaps | |
+| `baseline capture\|diff\|trend` | supported | versioned architectural metrics over time | trend reporting, not gating |
+| `retrieve` | experimental | typed knowledge queries over the model | |
+| `archetype` | experimental | evidence-based architectural archetype | |
+| `rename-class` / `chunk-file` | supported | word-boundary Java rename; split a large file for an agent | |
+| `cache status\|warm\|model\|clear` · `auth` · `telemetry` · `mcp` · `config` · `version` | supported | housekeeping | `activate` too |
+
+---
+
 ## What it does — and doesn't
 
 **ASK Engine reduces exploration cost.** It accelerates context acquisition and computes
@@ -182,6 +282,13 @@ Honest limits worth knowing before you rely on it:
 > **🎉 Early-adoption: Pro is currently unlocked for everyone.** Every install runs with
 > full Pro entitlements — no size gate, no key. The tiers below describe the model the
 > paywall will return to later.
+>
+> **What that means concretely.** `ask auth status` reports `"status": "unauthenticated"`
+> together with `"pro": true`, `"pro_reason": "early-adoption unlock"` — that combination is
+> expected, not a bug: you are unauthenticated *and* unlocked. When the unlock ends, gating
+> returns **by repo size and automation, never by command**: `posture`, `endpoints`,
+> `spring-audit` and `migrate-check` stay in the base tier at full output. Nothing you can run
+> today becomes a paid-only command tomorrow.
 
 **Gating is by repo size and automation — never by command.** Every command runs at full
 power on Free for small and mid-size repos; you upgrade when the work gets bigger or automated.
@@ -204,12 +311,18 @@ Full breakdown: [docs/PRODUCT_TIERS.md](docs/PRODUCT_TIERS.md).
 
 ```bash
 ask config              # version, config file path, telemetry status
-ask telemetry disable   # anonymous telemetry is on by default (opt-out)
+ask telemetry enable    # anonymous telemetry is OFF by default (opt-in)
 ```
 
-Telemetry collects version, OS, commands, flags, duration, repo-size range, and errors —
-**no source code, paths, secrets, or output**. Disable any time with
-`export SOURCECODE_TELEMETRY=0` (or `DO_NOT_TRACK=1`).
+**Nothing is collected or transmitted unless you turn telemetry on** — not on the first run,
+not in CI. If you do opt in, it collects version, OS, commands, flags, duration, repo-size
+range, and errors: **no source code, paths, secrets, or output**. Turn it off again with
+`ask telemetry disable`, `export SOURCECODE_TELEMETRY=0`, or `DO_NOT_TRACK=1`.
+
+> **Auditing someone else's code — regulated, client-owned or public-sector?** Nothing to do.
+> Telemetry was opt-out until 3.3.0; it is opt-in now (`docs/DEFECT-LEDGER.md` P-1), because a
+> default you must remember to disable is the wrong default for third-party code. An explicit
+> choice you made before is unchanged.
 
 **Custom security annotations.** Teach `endpoints`, `spring-audit`, and `explain` about
 project-specific authorization annotations via an optional `sourcecode.config.json` at the
@@ -241,3 +354,4 @@ Matching endpoints report `policy: "custom"` and drop out of the `no_security_si
 | [MANUAL-USUARIO.md](docs/MANUAL-USUARIO.md) | Guía de usuario en español |
 | [PRODUCT_IDENTITY.md](docs/PRODUCT_IDENTITY.md) | `ask` (command) vs `sourcecode` (package/alias) |
 | [privacy.md](docs/privacy.md) | Telemetry and data-handling policy |
+| [DEFECT-LEDGER.md](docs/DEFECT-LEDGER.md) | Every defect found in the field, its class, and which release closed it — published on purpose |