EFFECTIVE_MODEL: Requested opencode/glm-5.3-flash-free (max variant) was never active; this review ran on opencode/pulse-alpha-free (Pulse Alpha Free) with variant max, confirmed by the OpenCode TUI status bar readback ("... Pulse Alpha Free Default / OpenCode · max") in this session, which also matched an opencode models catalog query showing glm-5.3-flash-free as a registered, separate model ("GLM-5.3-Flash Free", variants low/high/max). This session was launched bare (no model flag) and OpenCode interactive mode offers no --model flag. runtimeMode: interactive OpenCode TUI worker session (Herdr terminal t7h, parent t75; Rex/hc hybrid per AGENTS.md). interactionMode: coordinator messages via hc send; all review tool calls ran directly in this session with no delegation. The mismatch was disclosed to t75 before dispatch; the review proceeded to schedule. Verdict written entirely by the confirmed running model above.

VERDICT: pass

BLOCKERS: none found.

Evidence trail (what the pass rests on):
- Contract verified first: plan hash resolves (repo copy of the frozen plan is byte-identical to /tmp/skill-lens-u8-final-review/plan-frozen.md, SHA-256 01f6f250..., differing only in the line-6 Reference substitution that commit a9d67b8 makes and the plan itself documents); HEAD is a9d67b8; working tree == tracked portion of the patch (git diff saved outside the repo is byte-identical, diff exit 0; new help.py = 325 lines, test_tui_help.py = 3110 lines as patched).
- No drift during review: SHA-256 spot check of 13 files (all 8 modified TUI sources' targets, tests/test_tui_help.py, tests/test_tui_home.py, tests/test_tui_read_only.py, the frozen plan, mutations/results.json) matches /tmp/skill-lens-u8-final-review/before.json exactly; the eight 4-byte blat artifacts were never touched.
- Framework semantics checked in the installed Textual 8.2.8 source, not assumed: first-pushed screens get no ScreenResume; pop_screen posts ScreenResume to the new top screen synchronously while the removed screen's removal is queued — the exact ordering hazard the hooks and tests guard.
- session.py, json_view.py and the controller are UNCHANGED by U8 (controller edits were never in scope per the plan; the diff backs this). Exact-origin/identity acceptance therefore rests on the same code that earlier phases already reviewed; U8 only adds callers.

Per-question results:
1) Help keyboard handling — no defect. F1 is an app-level priority bind that toggles only the exact active Help (app.py 229-232); q/r/? are explicitly rebound on HelpScreen because a modal stops ordinary non-priority app bindings, so q/r cannot bypass Help and ? cannot double-open it; printable keys while Home's filter Input has focus belong to the Input, and ? opens Help only outside it. Tests 200-254 (parametrized over all six), 427-512, 485-512, 604-640, 683-755 exercise each of these through real pilot presses.
2) Covered-worker ownership — no defect. open_help captures help_context() and calls origin.suspend_help() BEFORE push_screen (app.py 249-253) and itself starts no work; Help never computes, reads files, or dispatches. Each covered screen cancels ONLY its own token on suspend (why.py 352-362, diff.py 321-334, compare.py 373+) and refuses note_report_outcome/sync while suspended (why.py 559-560, diff.py 423-424, compare.py 456+); Home defers its why-queue via _suspended (home.py _drive). A report that lands behind Help is dropped by the covered screen and re-requested at most once on reveal (diff.py 358-363 re-request only when _report is None, owned request no longer active, and not failed; why.py 364-387 same via cache/owned guards). Mutation 04/05 prove these suspension gates fail-closed.
3) Refresh/reveal races — no defect. A capture published while Help is up does not touch covered screens: _update_ui routes HelpScreen to screen.update_session (app.py 593-597) and returns; reconciliation happens only on reveal via resume_help → sync_from_controller (why.py 583 closes a removed-detail Why itself with a notice before any restore; diff.py 347, compare.py match). Scroll/detail restore is same-generation-gated everywhere (why.py 510-522, diff.py 355-383, home _help_restore_pending), so a refresh behind Help can never resurrect old bytes (test 2076-2124; mutation 08 makes exactly this fail, test asserts old JSON document was replaced with the freshly serialized one).
4) Stale callbacks — no defect. close_help accepts only the exact expected instance (app.py 255-257), uses the tracked exact origin when it is still directly beneath, else falls back to the screen beneath (app.py 259-262). Every deferred helper re-checks is_mounted, app.screen is self, and _suspended before acting (why.py 501-546, diff.py 341-407, compare.py 392+); HelpScreen.update_session is inert before mount and after unmount (tested at 2739-2775, 3050-3059). Unmount clears owned tokens on wait, and the pre-mount/teardown test (2714-2736) walks every screen's bounded helpers.
5) Focus/scroll restoration — no defect. Every dismiss route restores exact Input text+cursor on Home (513-547), table focus and preview scroll (548-603), JSON scroll on all six screens (2558-2628), ordinary detail scroll on the five supported screens (2631-2706), with removal-fallback to the table for a deleted details pane (2992-3047) and truthful Help wording when a non-table control has focus (2806-2868, 3086-3110). Restore is skipped (not guessed) when the generation or widget state changed; Help body arrows only scroll when the body is focused, stated conditionally in the Help text itself.
6) Empty/error states — no defect. The A7 table distinguishes: successful-empty vs filtered-empty vs unreadable-root note (2267-2328, both sizes); zero detected agents vs zero registry definitions, with the registry-less Compare shortfall named exactly ("Need 2 more definitions"), (2331-2415); duplicate Compare choice rejected inline without a fault and explained in Help (2418-2445); Doctor empty vs info-only-healthy (2448-2506); single-copy/no-difference Diff (2509-2536) and not-found Diff keeping the model's own wording (2871-2910); no-session over Agents/Compare/Doctor honest and usable (380-426, 3062-3083); Why loading/unavailable agents notes (2913-2949). Accepted faults stay sticky behind Help: mutations 09/10 prove two different defects there (hiding the accepted-fault section; leaving Help's live session section stale after _update_ui stops refreshing it), and the read-only preservation test drives a sticky refresh fault behind Help, asserts the fault text in #help-session, and quits with exit code 3.
7) Mutation coverage — audited, not trusted. 12/12 categories map exactly to plan A8 (duplicate Help stacking; covered-key leakage; lost Input focus; ignored Diff suspension; ignored Compare suspension; stranded report after reveal; stale callback popping a newer screen; stale JSON after refresh; fault hidden while Help open; fault reset while Help open; banned import in help.py; write in help.py). I directly read the raw logs (not just summary.md) for cases 04, 06 and 08: each has control-before passing, a meaningful assertion failure at the intended test line, and control-after passing, and none is a syntax/import/collection failure. results.json records the planted diff for every case; the runner enforces all three conditions. The campaign ran on disposable copies outside the reviewed checkout.
8) Read-only regression: the patched test_tui_read_only.py adds help.py to the module inventory of the AST read-only guard and a full preservation traversal (Help over every screen, JSON toggles, refresh fault, sticky quit) with a byte-for-byte before/after tree-state comparison and 0o640 real targets.

SUGGESTIONS: (non-blocking)
- S1 tests/test_tui_help.py: several long user-story tests bundle 5-10 behaviors each (e.g. 1716-1925 and 2152-2259). If any grows flaky it will be costly to bisect; consider snapshots per behavior in a future low-risk task, not in this one.
- S2 Mutant runners reused two tests per category pair (cases 6 and 7 reuse test_help.../diff...). A second-order mutant that breaks both effects in one test could theoretically pass; extending the mutation set later to distinct assertion sites would close it. No current evidence this is hiding anything; control+failure+control held for all 12.
- S3 help.py's F1-description text describes filter typing behavior in prose (Home section). Any future change to Input key handling has to keep that prose in sync; there are content assertions for focus-naming but not for every wording line of the key map.
FILES_AT_RISK: skill_lens/tui/app.py (open_help/close_help/_update_ui Help routing — the seams every guarantee above depends on), skill_lens/tui/help.py (modal bindings and update_session guards), skill_lens/tui/why.py, skill_lens/tui/diff.py, skill_lens/tui/compare.py (cover/reveal hooks and deferred helpers), skill_lens/tui/home.py (suspend_help/_help_restore_pending), tests/test_tui_help.py, tests/test_tui_read_only.py. No file requires a change as a result of this review.
CONCEDED:
- C1 "Diff's native on_screen_suspend/on_screen_resume fire on ANY cover, not only Help" — withdrawn: in the current tree the only screen that can be pushed over an active Diff is Help (every other push originates from Home/Why below it, and close_diff/open_why pop rather than cover), so "any cover" and "Help cover" are the same set today; a future covering screen would be introduced together with its own plan.
- C2 "Help body arrow keys do nothing unless the body is focused — could mislead" — withdrawn: the Help text states the condition, and body is auto-focused at mount, so the ordinary path scrolls as promised.
- C3 "Home's suspend_help early-returns when already _suspended, which could skip re-capture of a legitimately re-suspended screen" — withdrawn: idempotency is the purpose (double suspend from the native event plus an explicit call), and the seeded-suspend test (2714-2736) pins exactly this safe behavior.
- C4 "Why's resume_help still runs its restore path after a removal-detected auto-close" — withdrawn after reading why.py 572-589: the removal branch calls _host.close_why(removal_notice=...) and returns BEFORE any restore; resume_detail returns on a True sync (384-385), so no stale restore is observable.
