feat: prepare v0.1.23 release (docs truth pass, ledger audit, changelog)

CHANGELOG.md

@@ -0,0 +1,82 @@
+# Changelog
+
+JevCompass publishes a focused changelog per release. Evidence boundaries and pilot
+limits live in [PILOT.md](PILOT.md); release-task receipts live in [TASKS.md](TASKS.md).
+Earlier releases predate this file and are summarized in
+[RELEASE_NOTES.md](RELEASE_NOTES.md) and the
+[GitHub releases](https://github.com/acidkill/JevCompass/releases) page.
+
+## 0.1.23 (2026-10-01)
+
+This release publishes the three explicit, nonblocking coding-decision commands that
+have been maturing on `main` since v0.1.22, together with the fifth bundled skill and
+the typed-decision cache. Nothing in this release gates commands, runs tools, changes
+permissions, or waives required validation, and no speed, quality, or cost benefit is
+claimed: the paired-trial evidence to date remains mixed or censored
+(see [PILOT.md](PILOT.md), VCR308/VCR312).
+
+### Added — user-facing
+
+- **Pretask strategy choice:** `jevcompass strategy choose --kind coding --signal …`
+  returns up to two reviewed strategies from allowlisted signals only; it accepts no
+  prompt, path, source, or log text. Verified local contract evidence
+  (`--contract-evidence consistent|conflicting|absent|partial|unknown`) and explicitly
+  resolved strategies (`--resolved-strategy`) route locally without a remote call.
+  Dependency-change inspection advice includes a short local migration checklist.
+- **Opt-in strategy hook:** `jevcompass install --strategy-advice` attaches conservative
+  pretask guidance to `UserPromptSubmit`; `--disable-strategy-advice` removes it.
+  Ordinary installs are unchanged and the hook has no evaluated coding benefit.
+- **Post-change test order:** `jevcompass tests discover` finds Python focused checks
+  from staged, unstaged, and untracked changes; `jevcompass tests rank` orders explicit
+  or discovered candidates. It prints an order and never executes tests; the supplied
+  required gate is returned unchanged. Optional verified metadata covers change
+  signals, per-candidate coverage and runtime buckets, and coverage targets. Bounded
+  `decision_reason` values distinguish local resolution, remote choice, and fallbacks.
+- **Failure triage:** `jevcompass triage --exit-code … --kind … --hypothesis …` ranks a
+  first diagnostic step for ambiguous failures using enums only. Verified observation
+  facts (`--import-observation`, `--assertion-observation`, `--timeout-observation`)
+  can resolve or constrain the choice locally and skip the remote call. Opt-in
+  `--rank-hypotheses` adds a bounded pairwise causal ordering. Output includes a fixed
+  `decision_reason` and per-step `selection_source` provenance.
+- **Caller-verified diagnostic costs:** `--diagnostic-cost HYPOTHESIS=low|medium|high|unknown`
+  passes fixed relative-effort tokens for supplied hypotheses only. Costs may guide
+  next-check ordering; they are never treated as causal-likelihood evidence and never
+  waive required checks.
+- **Typed-decision cache (opt-in):** `JEVCOMPASS_TYPED_DECISION_CACHE=1` reuses
+  validated strategy, test-order, and triage decisions for up to 24 hours with explicit
+  `cache_hit`/provenance labels; disabled by default.
+- **Fifth bundled skill:** the opt-in installer now ships
+  `jevcompass-coding-workflow` alongside `jevcompass-focused-tests`,
+  `jevcompass-regression-review`, `jevcompass-plan-implementation`, and
+  `jevcompass-plan-cutover`. Skills remain explicit opt-in and never install silently.
+- The `Changelog` project URL now points to this file.
+
+### Added — repository trial tooling (not in the wheel)
+
+- Paired-trial runners, supervisor triage bridges, case profiles, and immutable
+  fixtures for pretask, test-order, and triage comparisons, including equal-arm
+  mandatory-command exposure, safe shell-wrapper normalization, profile diagnostic
+  costs wired end to end through the bridge, and observation flags ordered before
+  `--rank-hypotheses` so intercepted commands stay bridge-counted.
+
+### Fixed
+
+- Triage separates causal hypotheses from diagnostic actions in pilot contracts.
+- Coding intent is preserved across mandatory validation commands.
+- Validation-timing status labels align with measured evidence.
+- Safe simple shell wrappers are normalized without granting compound-command gate
+  credit.
+
+### Verification
+
+- Full offline suite: 858 Python tests pass on Linux at this revision; the hosted CI
+  job runs the same suite on Python 3.11 for the release pull request and again inside
+  the publish workflow. Distribution contents are checked by
+  `tools/check_distribution.py` in the publish workflow.
+- Registry publication receipts (Trusted Publishing run, artifact digests, fresh
+  install check) are recorded in [TASKS.md](TASKS.md), [PILOT.md](PILOT.md),
+  [README.md](README.md), and [ROADMAP.md](ROADMAP.md) after publication.
+- Unchanged boundaries: advice stays nonblocking; remote ranking sees only allowlisted
+  coarse metadata; original failing exits and required validation are preserved;
+  macOS runtime and fresh Desktop prompt delivery remain unverified; no repeatable
+  speed or quality benefit is established.

PILOT.md

@@ -1856,3 +1856,12 @@ The bundled English coding workflow explains optional caller-verified relative c
 Future paired trials may declare caller-verified relative diagnostic costs in a case profile for supplied hypothesis IDs only. Both arms receive the same cost disclosure; the treatment arm is told costs may guide next-check ordering only, are never causal-likelihood evidence, and never waive required checks. Costs reach the remote decision only as fixed low/medium/high/unknown enum tokens through the validated bridge spec, the exact triage command, the intercepted shim request, and the production triage call, keeping transport bridge-counted. Cost-bearing requests are rejected by the bridge and shim unless they match the supervisor-validated profile. Profiles without costs keep identical prompts, commands, requests, and decision state. The supervisor triage command now orders observation flags before `--rank-hypotheses` to match the bridge command and shim, closing a silent-fallback path for exact rank-plus-observation requests in future trials; no historical trial is rescored or rerun.
 
 This is prospective trial instrumentation; no pair was run. Ten new wiring tests plus updated doubles pass, including parity and interception for rank-plus-observation commands, and the full 858-test suite passes locally with `git diff --check` clean. Hosted CI remains required before merge. No efficacy, delivery, or release claim is made.
+
+
+## VCR409 — Ledger numbering audit: no missing records
+
+A full audit accounted for every VCR identifier used by the acceptance and task ledgers. Twenty-seven identifiers — VCR278, VCR282, VCR300, VCR301, VCR306, VCR307, VCR309, VCR313, VCR316, VCR327, VCR328, VCR337, VCR339, VCR340, VCR357, VCR358, VCR361, VCR363, VCR364, VCR368, VCR369, VCR378, VCR385, VCR386, VCR388, VCR390 and VCR406 — appear in neither ledger, in no commit subject across all refs, and in no other tracked file; they are unused identifiers, not lost or withheld pilot records. VCR283 is recorded here only; the workflow-integration gap it announced was later closed by the phase-aware skill and verified-content refresh it described. A companion documentation pass restated the current capability set in README, ROADMAP, RELEASE_NOTES and a new CHANGELOG.md while preserving every unverified-benefit caveat. No frozen result, gate or score changed, and overall acceptance remains not established.
+
+## VCR410 — v0.1.23 release candidate
+
+Version 0.1.23 is a packaging release, not an acceptance result. It publishes the explicit strategy, test-order and triage commands, diagnostic-cost metadata, the typed-decision cache and five opt-in skills; per-change details and boundaries are in CHANGELOG.md. Pre-publication evidence at the release revision: 858 offline tests pass on Linux, `git diff --check` is clean, the exact-tree wheel and sdist pass `tools/check_distribution.py`, and the registry was rechecked (latest published 0.1.22; 0.1.23 unused). Hosted CI and the tag-gated publish workflow remain the merge and release gates. This release claims no speed, quality, cost, host-delivery or platform benefit; the frozen cohorts remain censored or mixed and fresh Desktop delivery stays unverified.