Date: 2026-10-16
Scope: review step R3 (post-phase-42-review-plan.md). The Phase 38 to 42
records, the README, TEST_PLAN.md and the handover are read against the code
at 6d8f7e4 and against R1’s numbers (TFP-207,
benchmarks/2026-10-15-tfp207-regression/).
Status: findings complete. Doc fixes D1 to D6 were approved by Emil and
made from Cowork on 2026-10-16, uncommitted, for Code to commit. The Phase 39
amendment was locked the same day. The first pass (F1 to F4, S1 to S5,
U1, U2) is carried over from post-phase-42-review-notes-2026-10-15.md,
which stays as the working notes.
main = origin/main = 6d8f7e4;57 passed (7.8m);166 inputs, 166 identical;analyse.mts: with-bugs +5 / −0, clean 0 / 0, and no status change on
any test present in both runs;evidence.mjs differs from TFP-188’s only in its run list;image-alt evidence is the same in both full reports, and
clean has none.outcome.md (doc fix, below as D1):
analyse.mts’s isLink pattern
misses the external-link title form link "Unsplash" to https://… on ….
The test-count changes since Gate B (+39 and +38) agree.076fba5 was committed 22 s into unit run 1. So runs 2 and 3, the
regression and the template check ran on 076fba5’s tree, and only
unit run 1 started on bfd900a’s. The difference is
benchmarks/ only, so no result changes.analysis.txt.| with-bugs Gate B | with-bugs R1 | clean Gate B | clean R1 | |
|---|---|---|---|---|
| Tests | 461 | 500 | 451 | 489 |
| Failed | 84 | 89 | 46 | 46 |
| Run (wall) | 1235 s | 1535 s | 1049 s | 1131 s |
| Test time (sum of tests) | 1127 s | 1399 s | n/a | 1057 s |
| Median test | 1.61 s | 1.92 s | n/a | 1.63 s |
| Time in failing tests | 410 s | 493 s | n/a | 189 s |
| TFP-205 link tests | — | 39 tests, 104 s | — | 38 tests, 67 s |
Score, Standard only: 30 found, 0 hinted, 7 missed. Gate B’s 33 needed Comprehensive security for 75, 80 and 82.
R1 shows the same elements still uncovered. Correct the sentence, and record that “every element in context” holds only for clean’s elements other than forms.
comprehensive-scenarios, performance-bands and a
Comprehensive security case. All three passed 3 of 3 in R1. The tests
that did fail once each were read-only-guard.test.ts:168 (TFP-199’s
beacon case) and cookie-scope.test.ts:90. The list is now history, not
the current flake set. TFP-208 takes all five, plus TFP-169 and the
regression web server timeout (below).The app’s own run timeout caps at 1,800 s too (calculateRunTimeout,
maxTimeout), though app runs use 4 workers. The Phase 42 record’s
TFP-205 entry should state the run-time cost. Phase 39 (amended) takes
the margin as its input.
analysis.txt).18b47d8.calculateWorkerCount), and
Comprehensive runs 1 off a local host (workerCountFor)./#/category/hand-tools. The page drew a
different window of products on each load: 1 to 9 at Gate B, 6 to 14 at
the inspection, fewer by the test.
next build and then a start, missed
its 120 s once.
| # | Fix | File |
|---|---|---|
| D1 | The three corrections above | benchmarks/2026-10-15-tfp207-regression/outcome.md |
| D2 | S1: Phase 41 complete 2026-10-13, Phase 42 complete 2026-10-15; Phase 39 entry pointing at the amendment | product-roadmap.md |
| D3 | F3, F5, F6, U3, U4 as correction notes dated 2026-10-16, without rewriting the history | Phase 42 record |
| D4 | S3, S6: a short fresh handover replacing the Phase 38 one (state, rules, next ticket) | docs/HANDOVER_TO_CLAUDE_CODE_2026-10-16-review.md, old one deleted in the same commit |
| D5 | F4, S2: header, overview, feature table (Phases 35 to 42), roadmap pointer; limitation 1 rewritten | TEST_PLAN.md |
| D6 | F1, F2, S4, S5, U1 (with both benchmark numbers), U2 | README front door, last, after D1 to D5 |
All six are made. D4’s new handover is
docs/HANDOVER_TO_CLAUDE_CODE_2026-10-16-review.md. The Phase 38 handover is
still on disk, for Code to delete (git rm) in the same commit. Code
commits these as docs commits; Cowork doesn’t commit.
post-phase-42-review-plan.md under R2. It takes R1’s flake
table, the two open questions above, and Emil’s two additions
(2026-10-16):
check:template-unchanged, on a nightly schedule
and on manual dispatch, not on every push (13 to 14 min);