feat(capture+review): scrub personal/proprietary identifiers from screenshots

[nicolasiscoding]

What

Makes DOM-scrubbing of personal/proprietary identifiers a built-in part of the capture flow, plus a pre-publish review check. Closes #1.

guidewright-capture

• New Principle 4 — never show a real account, ID, or org name.
• Scrub step added to the red-box per-step loop (run last, right before the shot, so a framework re-render can't restore the real text) and to the workflow checklist.
• New references/scrubbing.md: a parameterized, product-agnostic evaluate_script snippet (login email, org/workspace name, UUID/tenant id -> neutral placeholders like you@example.com / ACME Corp / masked all-zeros), with verify guidance. Placeholders/patterns are configurable per product, not hardcoded.

guidewright-review

• Pre-publish identifier-leak check in the walk step, the checklist, and lens 8 of the rubric: flag any screenshot/alt-text/page text exposing a real email, raw UUID, or non-generic org/customer name (Blocker/Major; checkable read-only).

Evals

• capture-scrubs-identifiers — a described flow seeded with a real-shaped email/UUID/workspace name; the authored page must use placeholders and plan the scrub.
• review-catches-leaked-identifiers — new evals/fixtures/leaky-howto.md with planted leaks; review must flag all three.
• Bundle now 5 evals; JSON validated. plugin 0.2.0.

Testing note

guidewright's evals are behavioral (one agent-run per eval, LLM-judged). They were authored in the existing evals.json format and dry-run-validated: the edited skills satisfy every new assertion, and the pre-change skills would fail them (so the evals actually guard the new behavior). Full harness run (spawns per-eval agents) should be run by a non-fork session or CI.

All planted identifiers in fixtures are fictional; no real accounts/orgs/ids appear anywhere in the diff.

[nicolasiscoding]

Eval run — all green ✅ (16/16 assertions)

Ran the full behavioral suite for real (per eval: a performer agent follows the edited skill → a strict LLM judge scores each assertion). Not a dry-run.

#	Eval	Skill	Result
0	howto-stays-recipe	capture	✅ 3/3
1	conceptual-becomes-explanation	capture	✅ 2/2
2	review-catches-planted-issues	review	✅ 4/4
3	capture-scrubs-identifiers (new)	capture	✅ 3/3
4	review-catches-leaked-identifiers (new)	review	✅ 4/4

New behavior confirmed:

• capture-scrubs-identifiers — a capture following the edited skill replaced the real login email / workspace name / raw UUID with neutral placeholders (you@example.com, ACME Corp, masked all-zeros id) and planned the DOM scrub as the last step before each shot — while still producing a clean action-only how-to.
• review-catches-leaked-identifiers — the review flagged the planted email, tenant UUID, and internal workspace name as Blocker/Major in a prioritized report.

No regression: the 3 pre-existing evals (recipe discipline, Diátaxis type-fit, planted-issue review) still pass with the edits in place.