Tags: PsychQuant/issue-driven-development

chore: marketplace.json 同步到 3.0.0

plugin.json 在 PR #327 裡升到 3.0.0，但這個 repo 自己的 marketplace.json 還停在
2.112.0 —— 跑 /plugin update 的使用者會看到舊版號，拿不到這次的 breaking 契約變更。
release-flow 規則講的正是這個尾巴：release 只做了一半，直到 marketplace 同步為止。

版號與 description 都從 plugin.json 取，避免兩邊各自維護又漂移。

fix: stop reconstructing perspective-writer's rules path; read and di…

…sclose status

R4 hardcoded `.claude/rules/correspondence-<person>.md` and only handed a path
over if that file existed. Upstream moved to a three-source resolution order
after 3.0.0, so in most current workspaces that path does not exist — nothing
gets handed over, upstream returns status=generic, and the reply posts
uncalibrated while the rules sit in a skill beside it.

Verified live, not hypothetical: in a workspace using skill-packaged rules the
path is absent and calibration silently does nothing.

- Hand over only `recipient`; let upstream resolve (needs perspective-writer
  >= 4.2.0, which added subject-based resolution for exactly this).
- Do NOT restate the resolution order here. A second copy in a second repo
  drifts, and cross-repo drift is caught by nothing.
- Pin MIN_PW_CONTRACT=4.2.0 with `sort -V`. Lexical comparison is wrong in both
  directions: 2.11.0 sorts before 2.9.2, and 4.10.0 sorts before 4.2.0.
- Read the returned status header and disclose status=generic. That draft reads
  normally and is indistinguishable from a calibrated one; the header is the
  only signal, so not consuming it means the failure mode does not exist.
- Version shortfall degrades, it does not refuse. Soft integration's posture is
  "works when absent"; refusing would cost the reply itself to save a
  calibration.

Honest limit recorded inline: status=generic cannot distinguish "no rules on
file yet" from "rules exist but were not reached" — upstream returns one value
for both (their #6 known residual). The disclosure is therefore neutral and does
not imply an error.

Refs #289

fix: stop reconstructing perspective-writer's rules path; read and di…

…sclose status

R4 hardcoded `.claude/rules/correspondence-<person>.md` and only handed a path
over if that file existed. Upstream moved to a three-source resolution order
after 3.0.0, so in most current workspaces that path does not exist — nothing
gets handed over, upstream returns status=generic, and the reply posts
uncalibrated while the rules sit in a skill beside it.

Verified live, not hypothetical: in a workspace using skill-packaged rules the
path is absent and calibration silently does nothing.

- Hand over only `recipient`; let upstream resolve (needs perspective-writer
  >= 4.2.0, which added subject-based resolution for exactly this).
- Do NOT restate the resolution order here. A second copy in a second repo
  drifts, and cross-repo drift is caught by nothing.
- Pin MIN_PW_CONTRACT=4.2.0 with `sort -V`. Lexical comparison is wrong in both
  directions: 2.11.0 sorts before 2.9.2, and 4.10.0 sorts before 4.2.0.
- Read the returned status header and disclose status=generic. That draft reads
  normally and is indistinguishable from a calibrated one; the header is the
  only signal, so not consuming it means the failure mode does not exist.
- Version shortfall degrades, it does not refuse. Soft integration's posture is
  "works when absent"; refusing would cost the reply itself to save a
  calibration.

Honest limit recorded inline: status=generic cannot distinguish "no rules on
file yet" from "rules exist but were not reached" — upstream returns one value
for both (their #6 known residual). The disclosure is therefore neutral and does
not imply an error.

Refs #289

fix(idd-comment,gh-egress): tier floor is a minimum, not a replacemen…

…t — verify round 1

6-AI verify round 1 (4 lens + Codex + DA) findings, all folded:

- HIGH (Codex + DA concur): SKILL R1 read "after confirmation dispatch at
  --scrub-attested warn" — a pinned level that would silently DOWNGRADE a
  third-party repo's ENFORCE tier. Rewrote as floor semantics: effective tier
  = max(repo-derived tier, warn); ENFORCE never downgrades, block-with-diff
  runs unchanged. Spec: warn-proceed scenario qualified to LIGHT/WARN repos +
  new enforce-preservation scenario.
- HIGH/MED (Codex + DA): layer-1 "already on same repo remote" premise was
  unguaranteed — a cross-repo comment URL is a new cross-audience disclosure.
  SKILL + spec: layer-1 exemption requires the URL to belong to the
  destination repo; cross-repo URLs resolve as layer 3 (marker records
  user-pasted) and carry the floor. New spec scenario.
- MED: rules/header reclassification — items 1-3 are level-independent
  zero-tolerance nets; item 4 is a level-dependent tier-floor attestation
  check (the blanket "LEVEL-INDEPENDENT"/"zero-tolerance" claims no longer
  held). gh-egress header now enumerates all 4 items.
- MED: honest raw-substring wording (two fixed-string token matches over raw
  prose, fences not stripped, NO backtick escape hatch — the entity-@ parity
  claim was inaccurate on that dimension) + accepted-friction LOCK test
  (fenced discussion body at light → 13).
- MED (DA N-2): SKILL Step 4 now takes the floor max() before attesting —
  the SKILL side is the primary gate for own-private, not just the wrapper
  bounce. MED (DA N-4): template↔wrapper token binding assertions (silently-
  disabled-net risk). Conjunction negative fixture (user-pasted without
  type=reply → dispatch) + stderr remediation assertion.
- LOW: grep -Fq -- hardening; "body needs no redaction" wording softened;
  design/tasks append-only + R1/R4 claims corrected to match reality.

Tests: gh-egress 66, idd-comment-reply 35, aggregator 40 suites 0 fail.

Refs #272

fix(idd-comment): reconcile reply-type contract per 6-AI verify round 1

Verify round 1 (4 lens + Codex + Devil's Advocate) findings, all folded:

- DA-1 (HIGH): verify-before-claim was issue-level (git log --grep "#N") but
  the obligation is per-point. Rewrote R2: the grep is an entry point; the
  found commit/PR must actually contain the change addressing THAT point, and
  finding any #N commit must not mark all points resolved. Locked by 2 new
  drift-guard assertions.
- Codex + Regression F1: --points-from "required" vs spec "when unspecified
  default to issue-body" was a genuine contradiction. Reworded spec so the
  three layers describe how the REQUIRED value resolves (issue-body is an
  explicit sentinel), not a default for an absent flag. Dropped --body from
  reply's 必填 column (Step 2 + spec make it optional).
- DA-2 (MEDIUM): bootstrap TaskCreate order (build_comment_body before
  resolve/verify) contradicted the pipeline "順序是契約". Reordered so
  resolve_points_source + verify_before_claim precede build_comment_body
  (= R3 anchoring) and annotated the dependency.
- Security F4 + DA-3 (MEDIUM): reply is the only type reproducing third-party
  verbatim. Added scrub-wins-over-verbatim clause + noted SCRUB_LEVEL is
  repo-visibility-keyed (own-repo → warn/light, never enforce) so layer-3
  pasted external content needs heightened self-review, not tier default.
- Test strengthening (Logic LOW-1 / Requirements F3 / Codex gaps): anchored
  the weak "reply" docs needle, added assertions for layer-1 URL / layer-3
  paste / per-point template elements / degrade post-anyway / scrub-wins.
  Suite 19 -> 27 assertions.
- Doc drift: design/proposal run.sh -> test.sh, suite count 39->40.

run-all-tests: 40 suites 0 fail. reply suite 27/27. spectra valid.

Refs #269

chore: release issue-driven-dev v2.94.0 (selective git auto-tag)

Ships #85 — idd-issue tags idd-{N}-baseline (rollback anchor), idd-verify
tags idd-{N}-verified on Aggregate PASS (review snapshot); config auto_tag
(default-on, opt-out), idempotent + graceful-skip. Minor bump for the new
capability. Bumps plugin.json + marketplace.json to 2.94.0; adds CHANGELOG
entry + README Version History row.

Refs #85