Ship the three v1.99.2 post-mortem rules into skills/sdlc/SKILL.md

[BaseInfinity]

Three rules were established by evidence during the v1.99.2 cycle and recorded in the post-mortem on #622. None of them are in the shipped skill, so no consumer receives them. skills/sdlc/SKILL.md ships (it is in package.json's files and appears in npm pack --dry-run); the post-mortem comment does not.

They were deliberately excluded from PR #664 because editing skills/sdlc/SKILL.md mid-merge would have changed the candidate tree and invalidated clearances already posted against it. That was the right call at the time and is the reason this is a follow-up rather than an oversight.

Rule 1 — A directive is not a verdict

Never post a CROSS-MODEL-CLEARANCE payload under a reviewer's name without a fresh verdict from that reviewer on that exact tree. A directive, ruling, or recommendation recorded during an earlier consult is not a verdict, however genuinely that reviewer said it.

Evidence: on PR #662 (2026-08-16) a clearance was posted reading reviewer: fable-5-via-advisor-high, confidence: 96, verdict: YES, bound to a tree that reviewer had never seen. What existed was its directive from the #661 consult specifying that change. The limitation was disclosed in the comment prose — and scripts/merge-pr.sh never reads prose. It parses the JSON payload and matches reviewer strings, so it cannot distinguish a verdict from a recollection, or a reviewer that ran from one that did not. It counted.

A guard that a cooperative party defeats by accident, while trying to be honest, is not a guard. This is the third live instance on #657, and the most valuable, because the first two were adversarial-shaped and this one had no adversary and no bug.

The honest degraded path when a seat cannot produce a fresh verdict: one seat plus --user-approved naming the directive. Weaker on paper, and true. Do not manufacture a second seat.

It held on first contact — #664 ran seat 2 as two fresh verdicts bound to named trees, NOT_CERTIFIED on d81a6499 then CERTIFIED on 6a2bbdeb, and the NOT_CERTIFIED caught a consumer-visible defect four blind seat-1 rounds had missed.

Rule 2 — Verify the artifact, not the config

A claim about what ships must be built from the shipped artifact, not from the input that is believed to produce it.

Evidence: the v1.99.2 CHANGELOG's "what reaches your install" blockquote was first built from package.json's files field and asserted as the shipped set. npm auto-includes README.md and package.json regardless of files, so the list was wrong. Sol caught it in round 6. Rebuilt from npm pack --dry-run, which is the artifact.

Same class, different surface: a config or environment edit reaches no reviewer, so "fixed" must mean an OBSERVED observable, never an edited file (#525).

Rule 3 — A VOLUNTEERED tag does not end the driver's judgement

The convergence rule says a VOLUNTEERED finding authorizes no additional round. It does not say the finding is wrong, or that the driver may stop thinking about it.

Evidence: during v1.99.2, a seat-2 finding tagged outside the card was the one that identified a consumer-visible defect (the CHANGELOG listing repo-local artifacts as if consumers received them) that four in-card rounds had missed. The correct handling is: do not spend a round on it, and do not discard it — file it, or fix it inside the current pass if it is cheap and in the same file.

The tag governs round budget, not truth. Conflating the two turns a budget rule into a licence to ignore evidence.

Acceptance criteria

1. All three rules present in skills/sdlc/SKILL.md, each stating its evidence, in the sections where the driver would actually be standing when the rule applies (Rule 1 at clearance-posting, Rule 2 at release-claim writing, Rule 3 at the convergence rule).
2. cowork/skills/sdlc/SKILL.md byte-identical — tests/test-cowork-drift.sh enforces this.
3. Rule 1 must not be phrased so that it reads as a mechanism. Enforcement needs a launcher-produced receipt (run id, resolved model, tree digest) and that is feat: make the second reviewer seat a per-repo declaration — Fable high or Opus 4.8 xhigh #657, still open. The skill states the discipline and names feat: make the second reviewer seat a per-repo declaration — Fable high or Opus 4.8 xhigh #657 as the missing enforcement; it must not imply the gate checks this today, because the gate is identity-blind.
4. Watch the skill's size. It is already ~42kB and skill listings are budget-clamped, so an oversized skill can degrade to name-only in routing (see Audit: is the SDLC harness a workflow written as a skill? (sequencing and state vs atomic capability) #665). Prefer tightening adjacent prose over net growth.

Milestone

v2.0.0. These are consumer-safety rules about how the harness certifies its own work, they reach a consumer install via a file that ships, and all three come from defects that actually occurred rather than from speculation.