feat(pr-workflow): add PR-audit, reasoning-audit, and diagnostic skills

[MajorLift]

Adds eleven skills. Each encodes a failure mode observed in real validation runs, stated concretely enough to judge the guidance against the mechanism rather than the assertion.

pr-workflow — eight, and testing — one

skill	the failure it encodes
falsifiers-first	Falsifiers taken from the author's sentences probe the mechanisms the author already phrased clearly. The probe lands on the most legible line rather than the weakest, the method mirrors the author's so agreement carries little information, and there is no bucket for effects nobody claimed.
unintended-breakage	A change breaks something it never mentions, and the diff looks intentional all the way through. Nine surfaces, each with a mechanical check; a candidate without a check is not reportable.
silent-failure	A mechanism fails and nothing says so. Detectability is a property separate from correctness and is almost never tested, so the method induces the failure and watches for a signal that never arrives.
unmeasured-join	A conclusion assembled from individually true facts, joined by a step nobody measured. The facts check out, so the join reads as prose and goes unchecked — and the conclusion is often correct, which is what lets it survive review.
instrument-check	awk -v escape-processes its assignment, so a mutation runner wrote /^[sS]{1,4096}$/u where /^[\s\S]{1,4096}$/u was requested, narrowing a regex meant to be widened. A different test failed than the one targeted, the suite ran in full, every guard passed, and the run reported a falsification for a line it never wrote. The artifact echoed the requested mutation, so the two could not disagree.
scope-of-search	A grep keyed on the variable name searchParams reported (no occurrence) directly above its own output showing five params.get('utm_*') reads. A negative is a fact about the pattern, not about the tree.
coverage-partition	"The test has power" is a boolean; power is a distribution. Three mutations against one suite: 4 of 7 cases guard the parameter strip, 2 guard signature verification, 1 guards the value format. Five pass with verification entirely disabled.
selection-audit	Two counts matched and were read as corroboration. They were different selection rules landing on the same number — one selecting changed test files, the other a set roughly twice as large.

| red-on-base | A test that passes on the branch proves the branch is green. Only one that fails on base and passes on branch proves the change is connected to the reported bug — and a base failure for the wrong reason (missing import, absent fixture, unrelated red) produces an identical exit code and proves nothing. Moved from #84; renamed from falsifying-test. |

coding — two

skill	the failure it encodes
distinguishing-observation	Instrumenting the mechanism you already suspect yields evidence consistent with your hypothesis and equally consistent with the ones you never wrote down. Names the pairs indistinguishable from outside — a swallowed error and an unreached path, a cache hit and a correct recomputation.
observability-gap	Debugging a path before establishing what signal exists on it. Distinguishes absent from suppressed, since a log filtered by level, sample rate or a flag is a different problem from one never written.

Both sit beside flaky-test-detection and pair with the debug orchestrator in #98.

The decision that needs a call

falsifiers-first and unintended-breakage read the PR description in opposite orders, which is the substance rather than a style choice:

skill	question	description
falsifiers-first	does the change do what it claims?	sealed until the hypotheses are fixed
unintended-breakage	what does it do that it was not meant to?	read first, to fix the envelope

Sealing prevents the claims from steering which mechanisms get probed. Reading first is required because out of scope is not a property of code — it is a relation between code and a stated intent, and you cannot detect a departure from an envelope you have not read. Each skill names the other and states why the order differs, so neither gets harmonised into the other later.

Three properties worth knowing before reading the diff

The unclaimed bucket	falsifiers-first sorts results into supported / contradicted / unclaimed. The third is the product, and a description-led run cannot produce it. It catches undersold changes as well as oversold ones.
The inverted verdict	In silent-failure, a green suite under an induced failure means the failure is undetectable — the opposite reading from a mutation probe, where green means the test is vacuous. The artifact has to say which reading applies, because the output is identical.
Ranking by distance to discovery	Not by severity. A small error nobody can detect outranks a large one that pages someone in a minute.

Reviewing this

Merge #84 first	Seven of these link evidence, which ships there and is not on main. The relative links resolve on merge and are dead if this lands first. Alternative: drop the links, add them in a follow-up.
It splits cleanly	Along the groups above — say so and I will split it.
Weakest guidance	falsifiers-first's phase A dominates cost on a large diff and the skill only says to scope by mechanism, it does not solve it. Sealing is imperfect — CI check names, branch names and commit messages leak the author's framing — so the skill says to record what leaked rather than claim it did not.

Test plan

• node .github/scripts/lint-skill-entry.mjs — 0 errors across 58 skills
• No wiki-link syntax, no bare lane identifiers in any description
• Every ## Related link resolves on this branch, except the six forward references to evidence
• Reviewer check: does each skill name a failure specific enough to test the guidance against, or does any read as generic advice