feat(qa,qa-only): add --evidence-per-finding evidence layout

[itstimwhite]

Why

The default flat `screenshots/issue-001-step-1.png` layout works well for 1-5 findings but gets noisy past that, and it's a chore to hand a single finding to a developer — you have to pluck their three screenshots out of a shared bucket and explain which lines of the report apply.

Companion PR #1483 adds `$B record` (video evidence at the browse layer). This PR adds the QA-side complement so that when an interactive bug is captured on video, the `.webm` lives next to the rest of the finding's evidence in its own folder.

The shape mirrors the report-folder pattern that ships in Vercel Labs' `agent-browser` `dogfood` skill — that skill exists as report-only; ours integrates the same evidence shape into both the report-only `/qa-only` and the fix-loop `/qa`, with our existing severity/health-score model on top.

What

Opt-in flag `--evidence-per-finding` (or natural-language: `evidence per finding`, `one folder per bug`). Writes one self-contained folder per finding under a per-run report dir:

```
.gstack/qa-reports/qa-report-{domain}-{date}/
├── REPORT.md
├── findings/
│ ├── 001-critical-checkout-500-on-submit/
│ │ ├── finding.md # severity, repro, env, expected/actual
│ │ ├── step-1.png
│ │ ├── step-2.png
│ │ ├── result.png
│ │ └── repro.webm # OPTIONAL — present iff `$B record` was active
│ └── 002-high-search-no-results/
│ └── ...
└── baseline.json
```

`finding.md` schema (defined in the shared methodology):

• Severity / Category / Page / Detected
• What's wrong (one paragraph)
• Repro steps (referencing the step-N.png files)
• Expected vs actual
• Environment (browser, viewport, auth)
• Evidence file index

When to use it (also defined in the methodology, so the LLM picks correctly):

• Run produces ≥5 findings
• Any critical or high severity finding
• An interactive bug has video evidence
• Findings handed off as Linear/Jira tickets (folder zips to one attachment)

When NOT to use it: quick smoke runs, 1-2 findings, regression-mode reruns. The default flat layout stays exactly as it was.

Implementation note (shape, not size)

Per gstack's prompt-size guidance, the structure and `finding.md` template go through the shared `generateQAMethodology()` resolver (loaded into both /qa and /qa-only via `{{QA_METHODOLOGY}}`). Each leaf `SKILL.md.tmpl` only adds one new Setup-table row and a one-line pointer in the Output Structure section. No copy-paste across leaves.

Commits are bisect-friendly per the contributor guide:

1. `feat(qa,qa-only): add --evidence-per-finding evidence layout` — resolver + both .tmpl source files (78 insertions).
2. `docs(qa,qa-only): regenerate SKILL.md` — `bun run gen:skill-docs --host all` output (146 insertions, generated).

Verified

• `bun test test/gen-skill-docs.test.ts test/skill-validation.test.ts` — 712 pass, 0 fail. No regressions in resolver or template validation.
• `bun run skill:check` — all 10 host-output freshness checks green (one pre-existing `claude/SKILL.md` missing-generated warning reproduces on `main` without these changes; unrelated).
• Spot-checked rendered output:
  • `qa/SKILL.md` and `qa-only/SKILL.md` both contain the new Setup row, the Document-phase Evidence layout section, and the Output Structure pointer.
  • The `{{QA_METHODOLOGY}}` block expands identically in both files.

Out of scope

• VERSION bump / CHANGELOG entry — left for the merge so the entry stays in your voice. The flag carries its own discoverability via the Setup table.
• Telemetry on which layout users pick. Worth measuring after this ships if you care.
• Auto-detection ("pick per-finding when ≥5 findings"). Kept explicit for now — the heuristic is documented in the methodology so the LLM can choose, but the parameter is opt-in.

Pairs with

PR #1483 (`feat(browse): add record command for video evidence of interactive bug repros`). The `record` primitive is what produces the `.webm` that lands in each finding folder. Each PR is reviewable independently; the QA flag works without `record` (the .webm just won't be there).