pi-lens 4.3.0 → 4.4.1

package/docs/pi-lens-monitor.md

@@ -4,10 +4,39 @@ Read a LIVE pi-lens session's logs and report what the numbers say, so the
 maintainer does not have to. The monitor observes; it never edits code,
 never restarts anything, and never touches the session it is reading.
 
-Read first: `AGENTS.md` (the "Recurring defect shapes" catalog, especially
-shape 41: a fixed bound reached at p50 is a design defect), then
-`docs/pi-lens-investigator.md` for the forensics conventions this role
-inherits. This contract adds the standing readout.
+Read first: the engineering principles (`docs/engineering-principles.md`),
+then `AGENTS.md` (the "Recurring defect shapes" catalog, especially shape 41:
+a fixed bound reached at p50 is a design defect), then
+`docs/pi-lens-subagent.md` and `docs/pi-lens-investigator.md` for the
+forensics conventions this role inherits. This contract adds the standing
+readout.
+
+## Memory samples
+
+memory_sample records include bounded attribution fields for diagnosing host
+memory growth:
+
+- process.heapUsedBytes is the reading at sample time;
+  process.heapSettledBytes is null until a major-GC performance entry, then is
+  the latest callback reading; process.heapSettledMajorGcCount makes its age
+  visible.
+- process.externalNonBufferBytes is externalBytes - arrayBuffersBytes.
+- subsystems.persistWorkers.reviewGraph and
+  subsystems.persistWorkers.projectSnapshot are the latest asynchronous
+  Worker#getHeapStatistics() readings, or null before a worker answers.
+- subsystems.treeSitter.treeCacheTotalBytes counts source bytes represented
+  by cached trees. Native/WASM growth is attributed by
+  process.externalNonBufferBytes; no tree-count estimate is emitted.
+- subsystems.wordIndex.wireBytes is the UTF-8 length of the word index's JSON,
+  taken natively while the snapshot persist encodes the body, and is null
+  until that index's current serialized form is persisted, and again once the
+  serialized form is released at `agent_settled` (the `word_index_memo_released`
+  row in latency.log records the release and its trigger).
+- samplerDurationMs measures the sampler's own record-assembly wall time,
+  excluding the surrounding turn.
+
+All fields are in the existing record and are O(1) reads or latest-value
+snapshots; no additional telemetry sink is created.
 
 ## Inputs
 
@@ -49,7 +78,47 @@ inherits. This contract adds the standing readout.
 6. **Backlogs**: `lsp_document_drift` rows by disposition, files affected,
    `driftAgeMs` p50/p95/max; `agent_end_deferred_mutation_drain` durations and
    coalesced path counts; `deferred_format_file` runs with `changed:true`
-   versus total.
+   versus total. A formatter that never settles is visible only as an
+   anti-join (#3828): a `deferred_format_post_exit_resync` row with
+   `outcome:"abandoned"` (beside the `hook-await-exceeded` degradation
+   `off_hook:deferred-format-post-exit-resync`) and no
+   `deferred_format_late_resync` row on the same `filePath` afterwards. Report
+   the count of such files. A late row names what happened to the file
+   (#3828 r3): `resynced` is the healthy end of the chain; `unheld`,
+   `no-service` and `vanished` had nothing to sync; `deferred` is still queued
+   for the next drift pass; `failed` did not land. The in-band
+   (`--immediate-format`) caller writes the same rows as `inband_format_late_resync`
+   (#3858) with the same outcomes. Both callers write one
+   `format_late_resync_chained` row (`metadata.which`: `inband` or `deferred`)
+   when the give-up chains the late resync (#3873), so a give-up is countable by
+   `filePath` even for a formatter that never settles or an Escape; a chained
+   file with no later `*_format_late_resync` row on the same `filePath` is the
+   anti-join to report.
+   **Session decisions** (#3873; each row is one lifecycle event, none is per
+   occurrence in a loop): `session_handoff_slot` (`op`: `stashed`, `replaced`,
+   `taken`, `key-mismatch-left` (once per stale slot), `forwarded`,
+   `unconsumed-at-exit`; `by`,
+   `reason`, `keyHash`, `storeNames`, `ageMs`), `session_handoff_adopt` (one
+   per primary start: `tried[]` with `source`, `found`, `version`, `ageMs`,
+   `storeNames`, and `chosen`), `session_store_action` (one per declared store
+   per primary start: `action` `adopt`/`reset`/`skip`, `payloadPresent`,
+   `itemsIn`/`itemsKept`/`itemsDropped`), `session_scope_transition` with the
+   `end` (a scope superseded without a shutdown) and `demote` (a start in a
+   replacement gap that is not the successor the shutdown named: a subagent
+   that binds in the gap is one, a real successor an interrupting reload
+   displaced is another; a subagent beside a live primary writes none)
+   transitions,
+   `session_end_fence_rollup` (one per primary shutdown: `sources[]` with
+   `guarded` and `dropped`, plus totals). Read `read_guard_branch_retained` with
+   `payloadReads`: `kept 0, dropped 0` is a missing payload when `payloadReads`
+   is `null`, an empty read set at `0`, and an ignored payload when it is above
+   0 beside a `payloadVersion` that is not the current one.
+   `session_start_total` carries `basis`, `gapMs` and `lineageMatch`;
+   `agent_nudge` carries `fileKeys` (hash8), `originSessionIds`, `scopeId` and
+   `queueEpoch`, so one touch delivered in three drains is three epochs;
+   a no-client `lsp_touch_file` carries `candidates[]` (`serverId`, `rooted`,
+   `clientFound`, `generation`) for the file's primary servers only, 8 at most;
+   an auxiliary server is listed only in `clientScope: "all"` touches, the one scope that resolves its root (the with-auxiliary scope reports it in `auxiliary_readiness`).
 7. **Timeouts**: `lsp_diagnostics_timeout`, `lsp_nav_request_timeout`,
    `lsp_client_wait_timeout` counts with `serverIds`/`source`.
 8. **Delta**: for each of the above, the change since the previous readout,
@@ -69,8 +138,7 @@ inherits. This contract adds the standing readout.
 
 - Premise first: before naming a constant or a seam, read the code that owns
   it (`clients/lsp/index.ts`, `clients/pipeline.ts`, `clients/runtime-agent-end.ts`,
-  `clients/lsp/document-drift.ts`). The 2026-09-09 readout mis-named a
-  suppression window as a trailing debounce; the correction cost a round.
+  `clients/lsp/document-drift.ts`).
 - Bounded output: the readout is one comment or one file. Never one line per
   row of the log. Quote at most three raw rows, each cut at 300 characters.
 - No repo edits, no Git commands, no restarts, no writes under `~/.pi-lens`.

package/docs/pi-lens-retro.md

@@ -6,9 +6,11 @@ pointer, or a deletion. Advice that lives only in a transcript is not an
 output of this role. Source of truth for every rule here is this file and
 `AGENTS.md`; `.claude/skills/retro/SKILL.md` only points here.
 
-Read first: `AGENTS.md` ("Recurring defect shapes", "Orchestration and
-delegated work", "Test requirements"), then `.claude/skills/merge-train/SKILL.md`
-("Common mistakes"), which is where mistake rows land.
+Read first: the engineering principles (`docs/engineering-principles.md`),
+then `AGENTS.md` ("Recurring defect shapes", "Orchestration and delegated
+work", "Test requirements"), then `docs/pi-lens-subagent.md` when delegated,
+then this contract and `docs/pi-lens-merge-policy.md` ("Common mistakes"),
+which is where mistake rows land.
 
 ## When it runs
 
@@ -68,7 +70,7 @@ is involved. Every finding cites the source it came from.
    decision where the line changed behaviour, or delete the line. A rule that
    was violated the day it was written is a mechanisation candidate, not a
    candidate for stronger wording.
-6. **Record.** One mistake row in `.claude/skills/merge-train/SKILL.md`
+6. **Record.** One mistake row in `docs/pi-lens-merge-policy.md`
    ("Common mistakes") per distinct mistake, with the fix that is now in place.
    Orchestrator-level lessons that are not repo rules go to the orchestrator's
    memory. Never a fourth home for rules.

package/docs/pi-lens-reviewer.md

@@ -1,71 +1,212 @@
 # Reviewer contract
 
+Read first: the engineering principles (`docs/engineering-principles.md`), then
+`AGENTS.md`, then `docs/pi-lens-subagent.md`, then this contract. Then read the
+issue's acceptance criteria, the full merge-base diff
+(`git diff origin/master...HEAD`), the PR body, and merge state. Principles §3
+("Reviewing and delegating") is the base of this role and is not repeated. Lane mechanics (checkout, `TMPDIR`, push forms, Git grants,
+summary shape) are `docs/pi-lens-subagent.md` "Orchestrator lane mechanics";
+a brief supplies only `lane=<name>`, scope, and the grant.
+
 ## Mission
 
-- Read the issue, full merge-base diff, `AGENTS.md`,
-  `docs/pi-lens-subagent.md`, the PR body, and merge state.
-- Keep the branch and worktree read-only.
-- Reproduce the claimed behavior through the production entry point.
-- Report only proven findings; do not repair the author's branch.
+- Break the PR before it merges. A finding a probe proves outranks ten you can
+  only argue. Reproduce claimed behaviour through the production entry point.
+- Keep the branch read-only and report only proven findings. Never repair,
+  push, commit, comment, or merge.
+- Write `REVIEW.md` at the worktree root, not only in the final answer
+  (#3261, #3264).
+- Facts about master come from a fetched `origin/master`, never the local
+  checkout (#2693).
 
-## Standard mechanics
+## Procedure
 
-- Write `REVIEW.md` as a file at the worktree root, not only in the final
-  answer. Two verifies this week (PR #3261 r3 and PR #3264 r3) delivered the
-  review only in the answer text.
-- When the fixer settled before its evidence pass, run the mutation table
-  yourself and say so.
+1. Merge state first: `gh pr view <N> --json mergeable,mergeStateStatus`, or
+   `git merge-tree --write-tree origin/master HEAD` when GitHub is flaky. A
+   conflicted PR is the top finding; report it immediately.
+2. Read the neighbourhood, not only the diff: every caller of what changed,
+   every callee it now reaches, every sibling seam that does the same job, and
+   every test double that depends on the changed shape (#2568, #2583, #2585).
+   Screen the diff against the `AGENTS.md` defect catalog.
+3. Verify the red-first claim: revert the source, keep the tests, rebuild, and
+   confirm the claimed failures (`AGENTS.md` covers the behaviour-preserving
+   exception). When the fixer settled before its evidence pass, run the
+   mutation table yourself and say so.
+4. Attack with probes and quote the output. Probe scripts live outside the
+   worktree (`<worktree>/../probes-<pr>` or the lane's `$TMPDIR`): an untracked `.mjs`
+   inside it reds `tests/scripts/lint-js.test.ts` (#2865). Attack inversions
+   (real failures downgraded, healthy paths narrowed); concurrency (shared
+   state, retained settled promises, check-then-act across an await); session
+   boundaries (once-only state after `resetDegradationLedger()` or
+   `session_start`, probed with the SAME cached object); cadence (cooldowns
+   against the caller's real retry interval, both directions); input channels,
+   trust boundaries, strict consumers, and old-record parsing; and doubles'
+   fidelity, including the same double in sibling test files.
+5. Run the targeted suites, every test file that references a touched symbol,
+   and the governance selection in `docs/pi-lens-subagent.md`, after
+   `npm run build`.
+6. Read CI once (`docs/pi-lens-subagent.md`), confirm Unit tests executed, and
+   read the log of every failing check to judge infra against code.
+7. Clean up: revert mutations, delete probes, and leave
+   `git status --porcelain` empty.
 
 ## Verification
 
-- Use `git diff origin/master...HEAD` or the merge-base equivalent.
-- For any change involving the pinned retired-synonym identifier population,
-  run the exact-pin sweeps on the MERGE of `origin/master` + head, not only on
-  the head. `tests/config/glossary-synonym-sweep.test.ts` (#3279) asserts the
-  live (term, file) population exactly in both directions; require same-PR
-  re-pinning from its `UNPINNED`/`STALE` output. The 2026-09-23 evidence is
-  two green PRs merging red (#3279's pins predated #3283, fixed on master by
-  #3288), plus #3284's own `path` count red (cue-vet 5→6, dart-analyze 6→4)
-  until a trailing re-pin.
-- Build and run the targeted and required governance suites.
-- Revert or neuter the source fix and verify the red-first test fails.
-- Mutate every new guard, filter, cap, fallback, and lifecycle path.
-- Probe inversions, concurrency, input channels, trust boundaries, strict
-  consumers, durable-record compatibility, and old-record parsing.
-- Repeat the pattern and population sweeps.
-- Check blast radius, bounded observability, changelog, commit, and PR-body
-  requirements.
+- **Name the behaviour population.** When a change replaces, moves, or widens
+  a lifecycle, dispatch, or ownership seam, run EVERY suite that exercises the
+  behaviour the seam governs and name the list; a curated file set cannot
+  clear what it omits (#3622).
+- **Generalization verdict.** When a mechanism, policy, guard, or optimisation
+  targets N of M members (registry entries, servers, tools, languages,
+  stores), name M, why the others are excluded, and their default, and end
+  with *widen in this PR*, *follow-up issue* (with its seam group), or *stay
+  specific* (with the reason). A missing verdict is a finding (#3622; the
+  optional ast-grep assist is #3684).
+- Review the pre-push self-scan with `npm run astgrep:self-scan`; `pr-preflight`
+  runs the same gate after its build.
+- Require the exact-pin rule in `docs/pi-lens-subagent.md` on the MERGE of
+  `origin/master` and the head.
+- Apply AGENTS.md's one mutation layer (#4005): required new-guard proof and
+  demonstrated correctness gaps gate the review. Spot-check at most one of the
+  fixer's hand mutations. There is no per-PR Stryker comment or `MUTATION`
+  line to read; the nightly Stryker report on master is exploratory and is not
+  a review input.
+- Repeat the pattern and population sweeps. Check blast radius, bounded
+  observability, changelog, commit, and PR-body requirements.
 - For LSP, dispatch, cache, runner, or tool changes, test one non-TypeScript
   registry entry through the same seam.
 - On a net-count fold, mutate every predicate the deleted sibling used to
-  back; the fold's own tests were written when two guards existed (#3064 F1,
-  #3065 F3, #3066 F3, #3068 F2).
+  back; the fold's own tests were written when two guards existed (#3064,
+  #3065, #3066, #3068).
+- Flag any new rule predicate added outside its owning domain module (#3781,
+  #3794, #3796). Review a behaviour-preserving move commit for caller-result
+  parity separately from any later behaviour change (#3817).
+
+## Standing probes
+
+Run every probe the diff can trip and say which ran and what each returned.
+
+- **Ladder and deletion sweep.** Every ask names the ladder rung it serves. An
+  ask that adds a guard names its recurrence; an ask that deletes a defensive
+  call has grepped every caller and test double first (#2568). An ask that
+  causes a needless fix round is a review defect.
+- **Duplication and over-build.** Re-implemented machinery (a second warn-once
+  latch, a private extension-to-language table, a hand-rolled walker) is a
+  finding even when SonarCloud is green. So is a new shared helper with
+  surviving siblings, unless the body carries the sibling list, the
+  unsafe-to-fold reason, and the issue link, and plumbing with no consumer,
+  unless the PR names its forcing function. Name the skipped rung; some seams
+  are wide on purpose (`HISTORY.md` "SDK-reuse boundaries").
+- **Model covers the change, not only TLC green (#3802).** When the diff
+  touches a file mapped in `formal/coverage-map.json`, require a `.tla`/`.cfg`
+  change under any one of the row's families or a `TLA+ unaffected: <family> —
+  <reason>` line for any one of them; rows of 4+ families only print a note,
+  except `index.ts`, which gates on the lifecycle hook handler a hunk lands in.
+  A green `TLA+ models` run over an unchanged model proves nothing about the
+  new code.
+- **Red-proof audit.** A claimed red without its quoted transcript is a finding
+  of its own; reproduce it (step 3).
+- **Quoted-evidence audit.** Diff every CI line the body quotes against the job
+  log on the exact head. A line the log never printed is an integrity finding,
+  reported first.
+- **Pushed observability.** The `Observability` answer names a phase or ledger
+  kind in a stream monitors read without asking (`logLatency`,
+  `logSessionStart`, the degradation ledger), and the diff contains that
+  literal. A pull-only surface is a gap (#2513, #2526). A new or replaced seam
+  also needs a success-path record.
+- **Record read-back (#3875).** Every new decision branch on a session,
+  lifecycle or delivery seam names its record (sink plus kind), cites an
+  existing one, or says `none: <reason>` naming each file. A test reads the
+  record back through the real seam (the diff's own, or the cited record's
+  existing one); `none:` has no record to read, so the reviewer checks the
+  reason instead. The lint is a prompt, never proof. Demotion: the first
+  refusal on a merged PR that its reviewer judges not a decision moves the
+  seam refusal to a `::notice::` advisory until a fixture pins that shape;
+  the orchestrator counts the ten-green promotion streak (`pr-metadata.yml`)
+  from each merged PR's `PR body (advisory)` row in `ci-verdict`
+  (recurrence: #3873).
+- **Changelog fragment.** Front matter `section:` is one of Added, Changed,
+  Deprecated, Removed, Fixed, or Security, and `audience:` is `user` or
+  `internal` (`user` = a pi-lens user or agent can observe it; `internal` = CI,
+  tests, `formal/`, contributor docs, orchestration, refactors), followed by
+  exactly one top-level entry. A missing or unknown `audience` is an error.
+  Bullet style and a bold or plain title are the author's choice
+  (`.changelog/README.md`). `CHANGELOG.md` changes only in the rollups
+  `npm run changelog:release` generates.
+- **Sort comparators.** Every new `.sort()` or `.toSorted()` has an explicit
+  comparator (SonarCloud S2871); an order that feeds an identity (dedupe key,
+  cache key, hash input) compares code units, not `localeCompare`.
+- **Flake shapes.** A new real spawn, elapsed-time assertion, raw timer wait,
+  or `vi.waitFor(` outside `vi.useFakeTimers()` needs a
+  `// flake-shape: <detector> — <reason>` header and `wallClockBudgetInclude`
+  membership (#2547). A pinned file whose count falls below its pin needs its
+  baseline tightened; a stale ceiling re-admits regrowth.
+- **Platform skips and session-start resets.** Apply the `AGENTS.md` test
+  screens and the `SessionStartClassification` invariant.
+
+## Verification rounds
 
-## Finding format
+On `VERIFY <head-sha>` with a claims list: fetch the head, rebuild, re-run YOUR
+original probes for every claimed fix (the fixer's tests are not proof), probe
+each claim's edge, re-run the targeted suites, and read CI on that head. Attack
+the round's changed lines as a fresh PR; fix rounds introduce defects at about
+the rate they remove them.
 
-Order findings: `CRITICAL`, `HIGH`, `MEDIUM`, `LOW`, `NITPICK`.
+- First spot-check one previous round's mutation on the new head (#2583).
+- When a round retunes a threshold, tier, or predicate, drive the boundary
+  input the new condition cannot separate through the real seam: a cure for
+  over-triggering tends to ship silent under-triggering (#2983).
+- "Passes locally now" does not prove a fix for a defect that involves a
+  deferred producer or another worker; it shows only non-reproduction (#2955).
+  Say so in the verdict.
+- A prescription you write carries its own sweep of sibling call sites, or is
+  marked "shape, not verified across callers" (#2642). A prescription that
+  narrows a guard names its residual family and measured incidence (#3155). A
+  fixer who proves your prescription insufficient with a red is right; verify
+  the override on its merits.
+- Judge every exemption a round adds (`DECLARED_EXCEPTIONS`,
+  `EXEMPT_SESSION_STATE_FILES`, a hook-await pin, a generation-guard exemption)
+  as silencing or registration, per entry, with the reason quoted (#2654).
+- Route the round per principles §3 and `docs/pi-lens-merge-policy.md`
+  "Round routing", and say it in the verdict:
+  "contract-only; merge on green" when every finding is a body claim, comment,
+  literal, changelog line, or a prescribed remedy with its quoted red. A
+  verdict, guard direction, lifecycle hook, or failsafe keeps the verify.
 
-Each actionable finding contains:
+## Materiality
 
-1. Severity and stable id.
-2. File/symbol anchor.
-3. Reproduction command or probe output.
-4. Expected and observed behavior.
-5. Root cause, cost, and concrete remedy.
-6. Issue-acceptance or repository-standard classification.
+- Do not report style preferences, hypothetical extensibility, minor
+  line-count savings, or anything lint, oxfmt, ast-grep, or the governance
+  sweeps enforce. Report the few materially useful findings per attack
+  dimension. A dramatically simpler seam is a named output, never a
+  fix-round demand.
+- Security-class findings (injection, path traversal, secrets, unsafe
+  deserialization, redaction, trust boundary) need a demonstrated exploit
+  through a real input path; theoretical DoS, regex DoS, log spoofing, and rate
+  limiting count only when the PR claims them. Put an undemonstrated one under
+  `Could not verify` with what would have been needed.
 
-Severity requires a reproduced failure. A high-severity hypothesis without a
-failure scenario is at most medium.
+## Findings and verdict
 
-## Verdict
+Classify every finding as **fold** or **file** with a one-word reason
+(`docs/pi-lens-subagent.md` "Follow-ups"). Order findings `CRITICAL`, `HIGH`,
+`MEDIUM`, `LOW`, `NITPICK`. Each carries a stable id, a file or symbol anchor,
+the reproduction command or probe output, expected and observed behaviour, the
+root cause, cost, and a concrete remedy, and whether it is an
+issue-acceptance or a repository-standard finding. Severity requires a
+reproduced failure; a high-severity hypothesis without a failure scenario is
+at most medium.
 
-Start with one verdict: `merge as-is`, `merge after fixes`, or `redesign`.
-Then include:
+Start with one verdict: `merge-ready`, `needs changes`, `redesign`, or
+`conflicted`. Then give spec-compliance and standards-compliance findings under
+separate headings, red-run verification, test totals, CI judgement, and
+merge-order interactions with other open PRs, plus:
 
 - `Could not verify`: every blocked or environment-limited check.
 - `Named output`: structural insight not closed by the probes.
-- `Disposition table`: each prior finding as `fixed`, `not fixed`, `new defect`,
-  or `withdrawn (reason)`.
-- Cleared categories and exact-head identity.
+- `Generalization verdict`: the population, excluded defaults, and verdict.
+- `Disposition table`: each prior finding as `fixed`, `not fixed`,
+  `new defect`, or `withdrawn (reason)`.
+- Cleared categories (one compact list) and exact-head identity.
 
-Use short, active, plain prose. Never merge, push, commit, or silently repair.
+Use short, active, plain prose.