chore: MAINT-307 resequence sorting deck + correct false perf-enforcement claims

[MP2EZ]

Closes MAINT-307

Two post-batch corrections, both "the record is wrong."

A — Sorting deck sequencing (FEAT-293 follow-up)

FEAT-293 promoted the control-sorting drill to a primary surface reachable from Home. The philosopher approved the rewritten card wording but flagged position separately: medical-test-anxiety was card 3 of 12, and on a wrong answer the screen announces "Not quite. How you respond to the anxiety is in your control…". For someone with a real pending medical result that lands ~30 seconds into a cold entry from Home, before they've learned what the drill is asking of them.

Moved to position 9, after several low-stakes cards (traffic, unanswered text, work effort, weather) have taught the in-control / not-in-control convention.

Sequencing only — no wording changed. A test asserts the wording is untouched, so a future edit can't smuggle a copy change in as a reorder. The other new assertions pin the property rather than the arrangement: sensitive cards must stay out of the opening five, so authoring a new scenario won't spuriously fail but a card drifting back toward the front will. event-interpretation is included in that set because it invites the user to supply "a difficult event" — the card most likely to be loaded with real material.

B — Correct two false perf-enforcement claims

The repo asserted performance controls that do not exist, which is worse than a documented gap: it stops anyone from building the real one.

• post-launch-monitoring-runbook.md §5 claimed "That gate already exists: the on-device Maestro flow enforcing <200ms / 60fps." Verified across all 8 flows in app/.maestro/ — not one ms or fps assertion, only timeout: failure ceilings. crisis-button-reachability.yaml:34-35 says so itself.
• The Performance regression CI job comment claimed "production budgets still enforced via perf:crisis/perf:breathing/perf:launch scripts." MAINT-166 PR 7 deleted those scripts.

Replaced with what's actually true — stated precisely rather than swapping one wrong claim for another:

Budget	Actually enforced by
Crisis detection <200ms	Strict CI gate — assessment-performance.test.ts, run by this very job
Crisis button <200ms	Coarse jest proxy — measures synthetic dispatch, not tap→render
Breathing 60fps	Nothing → INFRA-306
App launch <2s, transition <500ms	Nothing — hand-validated

Also records that performance-regression-reporter.js does not gate (non-strict unless PERF_REGRESSION_STRICT=true), so crisis_response_ms: 9.38 is a recorded baseline, not a threshold.

This weakens INFRA-282 AC5's deferral rationale, which leaned on that Maestro gate. The defer still stands on its other leg (no traffic pre-launch), and §5 now says so explicitly.

The matching .claude/CLAUDE.md Validation Matrix claim is corrected separately on _bare, where that file lives.

Verification

npm run precommit  exit 0 — safety 108, clinical 106, unit 592 (+5), privacy 51

Safety re-check: no safety-surface paths in the diff → Phase 2.5 correctly skipped, no simulator needed.

🤖 Generated with Claude Code