Repository navigation
Record #744's hand-copy guard lesson in mechanisms.md - #746
Merged
Merged
Conversation
A guard over a hand copy of a pipeline set went vacuous twice in one review: pinning copies missed an uncopied class (emoji, `dr★`), and a sweep accepting any refusal let an unrelated check answer (space, `dr x`). Amends the mutation-test field note with the fix that held, and names the interpunct as the class the predicate-driven sweep cannot reach, both in the note and in the test's comment. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
derek73
force-pushed
the
docs/hand-copy-guard-lesson
branch
from
October 19, 2026 18:49
7db3255 to
83f5102
Compare
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #746 +/- ##
=======================================
Coverage 99.05% 99.05%
=======================================
Files 47 47
Lines 4642 4642
=======================================
Hits 4598 4598
Misses 44 44 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
The pins hold the copied constants and stay green with the check gone; the text-with-an-east-asian-interpunct MUTATIONS row is what fails. Also cites the review by PR number and narrows 'would pass' to the case where a later check happens to refuse the test text. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
derek73
added a commit
that referenced
this pull request
Oct 19, 2026
…s.md, compact shipped measurements (#759) (#764) * Archive decisions/differential-ledger to tombstones and the retirement entry (#759) The ledgers #664 retired took 33,103 words across eleven arcs, every one superseded by the snapshot. Each arc is cut to a tombstone under decisions/README's rule 3 (date, issue, verdict, SUPERSEDED by 2026-10-13 #664); the retirement entry stands in full. The three passages other entries cite stay as one line each: the 'Donald mc' decline (decisions/suffix-acronym-collisions, decisions/v1-xfail-triage), the initials view (decisions/R4, the 2.3.0 release log) and the Policy-annotation Declined entry (decisions/3-0-reevaluations). The full text is pointed to at fe85bb3. mechanisms.md#RECORDED-ROSTERS said the entry keeps the ledger rosters' measurements; it now points to git through it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Tidy mechanisms.md: one paragraph per label, contracts without case history (#759) Every entry's template labels (Problem shape, Contract statement, How it works, Lives in, Reach for it when) now open their own paragraph. That is load-bearing and the preamble says so: test_doc_citations reads a Contract statement to the end of its paragraph, so in the one-paragraph layout every sentence of an entry was quotable. ONE-PREDICATE-PER-QUESTION (1,856 words) and RENDER-HONORS-THE-PARSE (2,771) are cut to their contract, a short How it works and pointers. The two sentences code comments quote from outside the old contract proper ("where the reader comes AFTER the decider, ..." in _group.py, "a token the parse never saw ..." in _render.py) move into the Contract statement paragraphs, so no comment in nameparser/ changes. The instance census and the case history are in git at fe85bb3 and in the decisions entries the rosters now name; the one piece no decisions entry held, the facade's spliced-field divergence under a caller's conjunction (#528, measured 2026-09-13 and re-run on this tree), moves to decisions/R3 as a dated entry. LEDGER-RULE-SEPARATION and CROSS-RULE-OUTCOME-PINS, both already "Retired with the ledgers" stubs living in nothing, are removed; the only references left are prose names in decisions/differential-ledger. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Compact shipped measurements and superseded passages in the largest decisions files (#759) A sweep under decisions/README's rule 3 over the 22 largest files after differential-ledger, every change checked shipped in v2.5.0 or earlier (the unshipped PRs are #737, #741-#746, #755, #756). Three triggers only: a whole entry headed SUPERSEDED by, cut to a tombstone; a superseded passage inside a live entry, cut to its decision and what replaced it; a measurement paragraph of a shipped change (what moves, cost, blast radius, gate and grid results), cut to one line naming population, counts, baseline and PR. Verdicts, reasons, Accepted costs, Declined/Excluded/Open blocks, measurements of declined or unshipped changes, harvested framing lines, and anything quoted or pointed at from elsewhere stay. RECOMPUTE recipes a release-log bullet rests on stay; others fold into the one line with their PR. The two #731 entries (decisions/M2, decisions/T4), reverted by PR #733 the same day, gain the SUPERSEDED by marker rule 1 asks for and become tombstones. About 13,400 words come out across the 22 files (M2 14,882 -> 12,190, R4 13,819 -> 12,877, the rest listed in the PR). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Fix the design-docs review's findings on the #759 trim - decisions/T4: the #731 tombstone had cut a gap and a declined option that predate #731 (ad486f6) and still stand; both are restored verbatim under the tombstone, re-measured on this tree. - decisions/differential-ledger: the tombstones keep the tooling and dormancy arcs' harvested framing lines (rule 3: "a tombstone included"); the `_check_tree` and DEFAULT_BASELINE Declined entries, about code and process that survive the ledgers, stand as written; the initials-view tombstone says the 2.3.0 release bullet's evidence is at fe85bb3. - decisions/H5: the population line no longer calls `毛 泽东 Dr.` a rules example; the #623 line keeps the corpus-and-case-text null result and the 500-parse characterization; three bold heads keep their closing period. - decisions/S2: the #614 line no longer classifies the stress grid's 8,472 movers, which were not classified; it gives the 12-parse slice that was. - decisions/S6: the frame totals decisions/S9 says "hold exactly" are back; the verification-round head keeps "AND ITS RECIPE". - decisions/R3: the facade spliced-field divergence moved from mechanisms.md is a continuation of the 2026-09-13 #528 entry rather than a second entry of that date, so the existing pointers to "the 2026-09-13 #528 bullet" resolve to one entry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Fix the second review's findings; record the compaction traps in rule 3 (#759) - decisions/differential-ledger: the initials-view tombstone gave the cluster sizes as 19 / 66 / 11 / 2, which is cluster (iii) split by vocabulary; the three clusters are 27 / 14 / 98. - decisions/T4: the restored gap paragraph was written under #731 and names "the committed grid", which the revert deleted; the lead now says so, points at PR #732, and scopes the re-measurement (the 108 and the readings hold on a rebuild; the 270 was not re-run). - decisions/H5: the #623 frames line said frames move "only" where a part holds a particle; the names without one moved too, and pay less. - decisions/README rule 3 gains the two traps the reviews found, so the next compaction meets them: a tombstone keeps what the entry recorded about the state before its change, and a one-line replacement keeps each count's own classification. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Scope T4's declined-option claim; fold the tombstone trap into rule 3's definition (#759) Third review round, on a4fed35: - decisions/T4: the lead scoped only the 270 as un-re-run; the paragraph's "closes both (0 field and 0 report differences)" is the same #731-era measurement. An approximation of the option re-run 2026-10-19 closes the fields but leaves 27 of 396 rebuilt pairs differing in their reports, so the lead now says the claim is #731's, unconfirmed since. - decisions/README rule 3: the new trap sentence changed what a tombstone keeps without amending the definition two sentences earlier, which still read as a complete list. The definition now says it, and the trap sentence keeps only the one-line-replacement half. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Records the lesson from #744's review in docs/design/mechanisms.md, as AGENTS.md's "write the check down where the next person will meet it" rule asks. It amends the existing Field notes bullet "Mutation-test a new guard before believing it" rather than adding a new one.
The lesson. A guard over a hand copy of a pipeline set (tools/vocab/build.py's check 2) went vacuous twice in one review:
dr★compiled and the test passed.dr xpassed because the phrase-title check refused it.The fix that held: drive the sweep from
_tokenize._ignorable(the tokenizer's own predicate), match check 2's own messages, and record the negative control.New in this PR: the sweep's limit. The design-docs review measured that hiding the CJK-interpunct check (rules/T3) still passes the sweep. The tokenizer divides at that interpunct by context in
_tokenize_region, not in_ignorable, What holds that check is itsMUTATIONSrow (text-with-an-east-asian-interpunct, input先·生); theEAST_ASIAN_RANGESandINTERPUNCTpins hold only the copies and stay green with the check gone. The note says so, and the test's comment now names the row instead of implying the sweep covers every class.Checks
uv run pytest tests/v2 -k "doc or design or tokenizer_ignores": 1085 passed🤖 Generated with Claude Code