feat(sdlc): make falsify-before-review an executable step (#590)

[BaseInfinity]

Ships the two reusable lessons from the #590 / PR #595 post-merge audit into the skill that installs into every consumer repo. They were living in one maintainer's private notes; that's the wrong home for a rule that applies to everyone.

The finding

Fable and Sol both audited the merged PR. Asked for the single biggest source of wasted rounds, Sol answered:

The builder hands work to the reviewer before trying to falsify it adversarially. That single failure produced nearly all the expensive rounds here.

Its evidence was specific: round 2's prompt asked the reviewer to check publication races and pid reuse that the builder had not tested first. Its phrase — "using review as the builder's test phase."

What ships

Step 1, falsify first. Not "be rigorous" — a checklist with a done condition. One row per factual claim and per risk named in the prompt, recording claim/risk | falsifier | route | evidence, routed through the file's existing EVAL / plain-assert / JUDGMENT-ONLY three-way call, with separate known-success and known-failure rows for every detector. Launch only when every row holds evidence or an explicit judgement-only disposition.

A direction demand, folded into the run-reviewer step. A prompt must ask for the smallest change that would satisfy the certify condition — and that direction is a proposal under the dialogue loop: implement it, dispute it with evidence, or route it to the brain if it changes the design.

How the reviewers changed it

This PR is its own argument, so the corrections are worth naming:

• Sol rejected the first draft as unexecutable exhortation. "Falsify your own work" is worthless if the reader can't tell when they're done. The checklist is its text.
• Both flagged that the draft contradicted this file's own scoped-TDD rule (RED only where a RED mutation is writable) — and this prose change is the counterexample. Routing through the three-way call resolves it.
• Sol caught the draft breaking this repo's claim rule, which is the sharper correction. It asserted "each one cost an architectural round" and "could have falsified in minutes" — neither measured, neither timed — and stated a causal collapse where the record supports only a sequence. Fixed to what the record shows. A file that tells readers not to ship unmeasured claims cannot ship one in its own rationale.
• Sol caught "implement verbatim" contradicting the dialogue loop's explicit right to dispute a finding.
• Fable caught the ordering: falsify-first was written after the step that launches the leg, so a reader executing top-to-bottom had already sent the review before reaching the rule governing what comes first.

Verification

Guards run before requesting the verdict — itself the lesson from #595, where certifying before the full suite let a CI-fix commit land post-clearance and cost a round: doc-consistency, cowork-drift (byte parity), hooks, workflow-triggers, session-load audit, skill-graduations, compliance, docs-usability, plugin — all pass.

Skill is 34,426 chars. The 20,000-byte ceiling was deliberately deleted in #489 as an unmeasured number, so there is none to violate — checked rather than assumed.

Two rounds. Sol: CERTIFIED 10/10, confidence 99.

[BaseInfinity]

CROSS-MODEL-CLEARANCE

Sol (GPT-5.6 Sol, high) — round 2 of 2. Round 1 returned three corrections and a smallest-passing-change; all applied with its own replacement text. Verbatim: "1. Yes. The certify condition is satisfied. 2. No remaining claim-rule violations. 3. VERDICT: CERTIFIED — 10/10."

The head SHA changed after certification only by dropping an unrelated local settings commit that had ridden in from local main. The certified diff is proven byte-identical across that rebase (sha1 aef90bc5118b0a0b before and after), so the certification covers this head's content exactly. CI validate green (3m49s) BEFORE this clearance was posted — the sequencing lesson from #595, where certifying first let a CI fix land afterwards and cost a re-bind round.

{
  "confidence": 99,
  "reviewer": "gpt-5.6-sol-high",
  "sha": "e387a9cb4e2da40b7c72f19f41b86bc6d97a20a4",
  "verdict": "YES"
}

[BaseInfinity]

CROSS-MODEL-CLEARANCE

Fable (high) — both reviewers shaped the final wording. The draft's three defects each got the reviewer's own replacement text: unexecutable exhortation became a checklist with a done condition, the contradiction with this file's scoped-TDD rule was resolved by routing through the existing three-way call, and the claim-rule violations in its own rationale were replaced with what the record shows. Fable's own finding was the ordering — falsify-first was written after the step that launches the leg, so a reader executing top-to-bottom had already sent the review.

Two rounds total, which is the cap this change itself prescribes for prose. Guards ran before the verdict was requested rather than after.

{
  "confidence": 96,
  "reviewer": "fable-high",
  "sha": "e387a9cb4e2da40b7c72f19f41b86bc6d97a20a4",
  "verdict": "YES"
}