give-me-tips + fallback-architect visible tip — QA evidence (2026-07-29)

Branch: feat/fallback-architect-tip-and-give-me-tips (commits cdc906a6c, 09cb34f10, 6ef405986, 8f8f5264d)

WHAT WAS TESTED

Unit (RED→GREEN, bun:test) — fallback-architect component: directive refusal-tip line, visible omo-fallback-architect:tip message (display:true) emitted only on refusal-driven fable5→fallback activation with the architect category active; tip-message.ts builder; skills-sync wiring for the new native skill. red-unit-tests.txt = failing BEFORE production code (right-reason assertion failures), green-unit-tests.txt = 63 pass / 0 fail after.
fallback-architect-e2e (real senpi 2026.7.28-2 + two-model mock provider, isolated sandbox) — e2e-red-old-bundle.txt: pre-change bundle FAILS only the new tip assertions (expectedTips 1 / observedTips 0 on refusal scenarios; negative scenarios pass). e2e-green-new-bundle.txt: rebuilt bundle, all 4 scenarios PASS (directive contains the refusal-tip content; transcript carries the display:true tip with fallback model id + give-me-tips; transient/architect-absent emit neither).
Interactive TUI proof (tmux, real pty) — tip-tui-driver.mjs --mode omo-fallback: live interactive senpi + refusal mock; the visible tip renders after refusal→fallback. omo-fallback.txt (plain pane), omo-fallback.ans (raw), omo-fallback.grid.json + omo-fallback.html (xterm-render triplet — assertions ran on the parsed cell grid), omo-fallback-receipt.json (tmux killed, sandbox removed, real auth.json untouched).

WHAT WAS OBSERVED

Refusal scenarios: hidden directive (with refusal-tip line) AND visible tip both in transcript; rendered on screen with the fallback model id (mock-weak) and the give-me-tips pointer.
Negative scenarios (transient fallback, architect category absent): no directive, no tip.
give-me-tips skill syncs via sync-skills.mjs into plugin/skills and is listed by install.mjs; skills-sync test updated (19→20 skills) and green.

WHY IT IS ENOUGH

The component's activation path (detection + gate) is untouched and pinned by the e2e negative scenarios; the additive tip is proven at the transcript seam (e2e) AND the rendered surface (tmux + cell grid). Unit RED→GREEN covers the builder/gating logic. Remaining regression risk: renderer styling only (cosmetic).

ISOLATION NOTE

The e2e whole-dir digest guard reported contamination: a live operator session (and its two implementer child sessions) was concurrently writing ~/.senpi/agent during the run. e2e-green-attribution.txt lists every changed path — all are live-session transcripts/logs (sessions/*.jsonl, settings.json, debug logs); auth.json is sha256-identical before/after. The QA children wrote only to their /tmp sandbox agent dirs (removed; receipts in omo-fallback-receipt.json).

WHAT WAS OMITTED

No secrets captured: mock provider uses a file:// baseUrl + inline "mock" key; sandbox settings contain no credentials; live-session transcript paths are listed by name only, contents untouched.

POST-MERGE (origin/dev 0dbf644f9) RE-QA

After merging origin/dev (bundle conflict resolved by regeneration): bun test from root 412 pass / 0 fail (bun-test-post-merge.txt); fallback-architect-e2e overall PASS incl. isolation guard (e2e-green-post-merge.txt).