Repository navigation
Description
Why
In the month to 2026-09-28, src grew from 62k to 118k lines and from 66 to 123 modules. That came from 685 commits, about 24 a day. src was 10.6k lines at the end of June.
The architecture at package level still holds, and it is machine-enforced: _LAYERS has no upward-import exemptions, and ADRs 1–5 are tied to guard tests. Inside modules, though, structure has decayed:
- God modules:
from_model.pyis 10.2k lines anddrawing.py6k. Seven modules take 64% ofsrccommits. - Very long functions: 27 functions are over 400 lines, and 147 take 8 or more parameters.
- Closures used as classes: loop state is captured through default arguments instead of a proper type.
- Duplicated helpers across the linting and contract modules.
- Untyped side-channel state stored on helper-library objects.
- Comments that record history instead of explaining the current code.
- A module map in the docs that no longer matches the code.
This review was read-only and done at main @ 71b8c94. All figures come from AST, grep and git scripts; each sub-issue carries its own evidence.
Sub-issues (suggested order)
Structure: highest leverage
- Code health: introduce BuildOptions to collapse the 51-parameter build signature copied across six front doors #1927
BuildOptions: collapse the 51-parameter signature copied across six front doors - Code health: shrink the Drawing god class (extract PDF text layer, lint orchestration, intent drain) #1928 Shrink the
Drawinggod class (PDF text Extract PDF semantic text layer from Drawing (#1928) #1966, lint orchestration Extract Drawing lint orchestration (#1928) #1968, intent drain Extract deferred intent drain from Drawing (#1928) #1973, evidence Extract Drawing evidence and suppression reporting #1998, intent/export/state/table owners Extract Drawing intent, export, and table owners #2009, and diagnostics/edit owners Complete Drawing facade extraction below issue limits #2012 extracted;drawing.pyis 1,940 lines with a longest method of 149 lines; independent review and CI passed) - Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 Placement-candidate type and function/closure limits (merged in Converge placement callback state and guard source shape (#1939) #2103; six named functions are each under 200 lines and the nested-default guard passes)
- Code health: split annotations/from_model.py (10k lines) per feature family #1929 Split
from_model.pyper feature family (completed in merged Refactor annotation owners and share build-local evidence (#1939) #2104; eachannotations/source module is at or below 2,500 lines, with the source-size guard and independent review passing)
Smells
- Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 Replace the
_dw_*setattr side-channel with typed state (completed in merged Consolidate typed evidence and suite debt (#1939) #2105; zero_dw_source matches, reintroduction guard and current-head CI passed) - Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 Consolidate genuinely equivalent contract and lint helpers (completed in Consolidate equivalent lint helpers and evidence tests (#1932, #1935) #2107; the accepted one-owner-per-equivalent-behavior criterion and source-deletion target are met)
- Code health: make compat shims warn and stop the suite importing through them #1936 Compat shims warn and the suite imports canonical owners (merged in Warn on compatibility shims and migrate suite imports (#1936) #1954)
Structure and docs (maintainer decisions)
- Code health: rationalise rank-0 module sprawl and make module docs track _LAYERS #1933 Rationalise rank-0 module sprawl; make the module docs track
_LAYERS(location, obligation, and leader-policy ownership slices merged in Fold location predicates into measurement support (#1933) #1972, Fold obligation policy into layout (#1933) #1976, and Move leader region policy into annotation layout ownership #1979; rank-0 and docs exit criteria remain) - Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 Hole/feature leader convergence is deferred until after this epic by maintainer decision. Its current paths are documented; the issue remains open outside this goal and does not block Epic: code health after the Aug–Sep growth (structure, smells, tests, performance) #1939 completion.
- Move contributor instructions from CLAUDE.md to AGENTS.md (and slim them) #1938 Move contributor instructions from CLAUDE.md to AGENTS.md and slim them (merged in Move contributor instructions from CLAUDE.md to AGENTS.md #1940)
- Code health: strip history/review-round citations from src comments #1934 Strip history and review-round citations from
srccomments (38-file first sweep merged in Remove review provenance from source comments (#1934) #1987; scanned review-round/Codex patterns are gone, but the issue-reference halving target remains)
Tests
- Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 Shrink issue-named tests, audit clone pairs and share useful fixtures (completed through Consolidate equivalent lint helpers and evidence tests (#1932, #1935) #2107: issue modules 151→142, diagnostic two-member groups 59→24; project-level covered fast-tier time/coverage acceptance holds with the 29-baseline-exclusion caveat recorded on Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935)
- Code health: audit and remove proven dead code paths #2106 Audit possible dead code with Vulture and remove only proven unused paths (completed in merged Remove proven dead private drawing helpers #2108; 51 net source lines removed, false positives documented, 17 successful CI checks)
Performance. On simple parts about 90% of build time goes to OCC work on annotation, text and title-block geometry, not on HLR or recognition.
- Perf: stop scale ladders rebuilding when no smaller scale can clear the blockers (CTC-02 >15 min) #1941 Automatic scale budget skips a provably impossible A0/1.0 build; the CTC-02 slow budget and parity gates passed (merged in Bound automatic scale retries by physical principal-view fit (#1941) #1950). The remaining explicit A2/0.2 ladder is tracked in Perf: bound explicit CTC-02 A2/0.2 scale fallback without suppressing recovery #1949.
- Perf: bound explicit CTC-02 A2/0.2 scale fallback without suppressing recovery #1949 Explicit CTC-02 A2/0.2 fallback stops on a proved short off-axis span (merged in Bound explicit fallback on proved short off-axis span (#1949) #1965; 9 to 6 full builds, 143.21 to 119.69 s, 14 live signatures identical).
- Perf: build the title block once per build (built 4x per assemble, 13 s on grm03) #1942 The title block is built about 4 times per build (13 s on grm03). (merged in Perf: cache title-block construction across build retries (#1942) #1947)
- Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 — separate open performance follow-up; four-workload measured acceptance remains unmet (maintainer exception).
- Perf: probe placement candidates with a cheap label model; reuse geometry across repack passes #1944 Conservative Dimension ink probe merged in Probe dimension ink before building candidate helpers #1983 (GRM03 constructors 152→27; crowded five-hole box 129→81). Conservative bound, live parity, and paired timing gates passed. Repack reuse was measured below 1% and declined; the issue’s measured acceptance is met.
- Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 — separate open performance follow-up; build-attempt counted-call acceptance remains unmet (maintainer exception).
- Perf: iso-growth search re-projects per probe — use one projection + similarity? (ADR 2 question) #1946 Exact iso similarity replaces repeated clearance probes (merged in Use exact iso similarity for clearance probes (#1946) #1953; 17 live signatures and ADR 2 differential proof passed).
Build times measured on main: 4–6 s warm for trivial parts, about 39 s for grm03. -m smoke took about 100 s, where the docs say 30 s.
Exit-criteria coverage to track
At main @ a6ae6de, ten src modules exceed 3,000 lines. #1928 and #1929 explicitly target drawing.py and annotations/from_model.py. Other issues touch some of the remaining files, but do not commit to reducing their module size:
- Track the other eight modules with explicit per-module worklist and acceptance in Code health: shrink the eight remaining source modules over 3,000 lines #1955. The 3,000-line limit is not yet met;
sheet_emit.pyreached 2,959 lines in Extract sheet object source helpers (#1955) #1977 andpmi.pyreached 2,910 lines in Extract PMI support blockers below module limit (#1955) #2003 after its Split audit comparison and PMI census into focused stages #1994 regression;model/ir.pyfell below the limit to 2,922 in Extract foundational IR records into focused rank-zero owner #1993, whileevaluation/step_analysis.pyis 3,760 after Extract chamfer and fillet STEP evidence #2016.annotations/holes.pyreached 2,969 in Extract off-axis hole location pass (#1955) #1990.sheet.pyfell below the limit to 2,922 in Extract Sheet identity and layout-control owners (#1955) #1992 and Extract Sheet view declarations and meet module limit #1996. A fresh census onmain@21e5cdc5has one module over 3,000 lines:annotations/from_model.py5,987 at that census; Code health: split annotations/from_model.py (10k lines) per feature family #1929 subsequently closed in Refactor annotation owners and share build-local evidence (#1939) #2104 with eachannotations/source module at or below 2,500 lines.evaluation/step_analysis.pyfell to 2,717 after the Extract flat and polygonal STEP evidence #2018/Extract hole-family evidence from STEP evaluator #2031 STEP stack andannotations/leaders.pyto 2,758 after the Extract greedy feature leader selection phases (#1930) #2027/Extract settled feature-leader fixed ink lowering owner #2036 leader stack; all Code health: shrink the eight remaining source modules over 3,000 lines #1955 worklist modules are now within the limit.annotations/_common.pyreturned below the limit at 2,971 in Extract strip postsolve owner from shared annotations #2030 after the Split shared strip placement into explicit phases #2028 strip-phase extraction; Code health: shrink the eight remaining source modules over 3,000 lines #1955 also tracks the oversized leader module.builder.pyfell to 2,696 in Extract builder attempts and automatic trials below module size limit #2007 anddrawing.pyto 1,940 in Complete Drawing facade extraction below issue limits #2012, closing Code health: shrink the Drawing god class (extract PDF text layer, lint orchestration, intent drain) #1928. - Track the function-length exit criterion: Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 owns its six named functions and nested
greedy; Code health: bring the remaining 33 long functions under 300 lines #1956 owns the other long functions. A fresh AST census onmain@21e5cdc5finds 19 functions over 300 lines after Extract greedy feature leader selection phases (#1930) #2027 reduced nested leadergreedyto 191; Bound shared corridor solve by extraction (#1956) #2023 and Split shared strip placement into explicit phases #2028 reducedsolve_corridorto 287 andplace_strip_candidatesto 121;build_oncefell to 288 in Extract Drawing intent, export, and table owners #2009. The hole-outcome function fell from 372 to 214 lines in Extract hole-source ownership join (#1956) #1986,render_slotsfrom 485 to 19 in Extract slot rendering into focused annotation owner #1995,render_step_lengthsin Extract turned step-length annotation rendering #2002, and_assembleto 296 in Extract builder attempts and automatic trials below module size limit #2007;compare_assessmentsand_extract_pmi_censusalso fell below 300 in Split audit comparison and PMI census into focused stages #1994. The overall 300-line limit is not yet met. - Link the performance sub-issues when profiling is complete, so that part of the epic has tracked work and an exit check (Perf: stop scale ladders rebuilding when no smaller scale can clear the blockers (CTC-02 >15 min) #1941–Perf: iso-growth search re-projects per probe — use one projection + similarity? (ADR 2 question) #1946).
Ground rules
- Preserve drawing behaviour in every refactor. The golden corpus and script↔CLI parity tests are the oracle; never re-bless them to make a refactor pass. The maintainer authorised only two narrow behaviour exceptions: Code health: make compat shims warn and stop the suite importing through them #1936 deprecation warnings, and Perf: stop scale ladders rebuilding when no smaller scale can clear the blockers (CTC-02 >15 min) #1941/Perf: bound explicit CTC-02 A2/0.2 scale fallback without suppressing recovery #1949 bounded honest failure when no smaller scale can clear blockers. Successful drawings must remain unchanged.
- Each extraction updates
_LAYERS,docs/architecture.mdand the agent guide in the same PR. - Code health: rationalise rank-0 module sprawl and make module docs track _LAYERS #1933 and Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 may touch ADR wording. They need maintainer sign-off before any ADR text is written.
Agreed execution decisions (2026-09-28)
- Base new work on main after Move contributor instructions from CLAUDE.md to AGENTS.md #1940 (
3cb98102). Use isolated worktrees, focused PRs, and a fresh independent agent review for each PR. - Merge eligible PRs after required local and CI gates pass and blocking review findings are resolved. Do not merge ADR text without maintainer sign-off on the specific wording.
- For Code health: rationalise rank-0 module sprawl and make module docs track _LAYERS #1933, fold small modules into natural consumers and group cohesive remainder. For Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937, converge leader engines in small verified PRs only if investigation finds no hole-specific reason to keep them separate.
- For Perf: iso-growth search re-projects per probe — use one projection + similarity? (ADR 2 question) #1946, replace repeated projections only if equivalence is proved and ADR 2 remains satisfied.
- Profile reproducibly on the execution machine. The performance work is complete when each issue's measured acceptance holds and the benchmark set has no regression; no fixed overall speedup target is required.
- Treat the approximately 3,000-line module and 300-line function limits as required for closing this epic.
Suggested exit criteria
- No
srcmodule over about 3k lines. - No function over about 300 lines.
- No closure capturing five or more default arguments.
- Rank 0 in
_LAYERSmeans "imports nothing in the package".
Active run status (2026-09-29)
-
The reviewed eight-PR stack Move deferred intent record to routing owner #2020–Extract compiled height-ladder corridor registration #2043 merged at
e801f6f9; Consolidate Sheet identity invalidation contracts (#1935) #2052 merged at0e64c646. The native 15-PR stack Extract feature leader trace and fixed ink owners (#1930) #2045–refactor: extract finished annotation layout comparison phase #2063 passed current-head required CI and independent reviews, then merged atomically ata73a0adf. The separately reviewed slow-canary correction Correct three pre-existing slow canaries #2065 merged at6e645c2f. No ADR text changed. -
Independently reviewed cleanup PRs Share exact recogniser capability key checks (#1932) #2053 and Consolidate pocket and slot pattern edit contracts (#1935) #2055 passed focused/static checks and required CI, then merged at
78dc0624and5d4f91b5. The separately approved Y-axis behavior-fix PR Preserve Y-axis step detail clearance during layout #2044 passed required CI and fresh review, then auto-merged atda37a4bc. Five stale fast-tier canaries reproduced on unchangedmain; reviewed test-only correction test: restore five fast-tier canaries after owner and layout changes #2066 passed required CI and auto-merged at63cb9fe5. Rootmainis pulled through that commit. -
Stack #2088 publishes 17 independently reviewed refactor slices as PRs Extract gusset pattern lowering stage #2067–test: require architecture map to match ranked source modules #2079 and Extract ordered model detection through typed run stages #2081–Extract slot callback state into typed owners #2084, based on
mainat63cb9fe5. PR test: track hole-table provenance reads across owner split #2080 was closed after its provenance guard moved into Extract hole table evidence reconciliation phases #2077 so intermediate heads have the correct test map. The integrated refactor source is byte-identical to the previously parity-checked tip: static and 110 focused architecture/compiled-plan tests, and 14 CAD plus three STEP exact drawing/lint/decision/trace parity checks. Its local census is zero source modules over 3,000 lines and zero functions over 300. The rebuilt tip passed static and 72 focused boundary/tier tests. Required CI is running on the rewritten heads; Extract ordered model detection through typed run stages #2081 was reopened to rerun classification after its first event used the deleted former test: track hole-table provenance reads across owner split #2080 base SHA. No open stack PR is counted as merged completion. -
Stack PR perf: reuse physical lint within finished build policy (#1945) #2085 remains a draft Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 lint-reuse slice. Its original implementation passed 219 focused tests and exact 17-case CAD/STEP signature parity but showed a 0.8% GRM03 slowdown across ten pairs. A selective-scope revision passed fresh independent review and focused/static checks. New ten-pair default-path medians are essentially neutral (GRM03 +0.18%, blind-hole STEP indistinguishable); six-pair explicit and compare Box medians are slightly faster. The no-regression gate and current-head CI remain open. Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943's bbox/deepcopy investigation found no safe patch meeting parity and timing; counts and discarded experiments are in Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943.
-
Code health: rationalise rank-0 module sprawl and make module docs track _LAYERS #1933 has five reviewed strict-leaf slices reducing rank-0 package importers from 19 to zero. Their post-stack integration passed static, 46 focused architecture tests, and exact 17-case CAD/STEP annotation/build-issue/page/view signature parity; the final guard was mutation-checked across runtime, type-only and lazy imports. Follow-on PR #2089 is open against the stack and awaiting CI. Code health: strip history/review-round citations from src comments #1934 has four reviewed comment-only slices reducing tracked Python comment issue references from 983 to 462, below the halving threshold with exact AST/non-comment-token parity; its rebased PR #2087 is awaiting CI. Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 has typed dimension-ink candidate and producer-rider slices, independently reviewed including a load-bearing rebuild guard; the integration matched 17 CAD/STEP drawing/lint/decision/trace signatures exactly. Follow-on PR #2090 is open and awaiting CI. Other
_dw_*families remain. None of these local/open results are counted as merged completion. -
Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 full leader convergence remains open. Sparse side/plan holes already enter the shared feature queue; front ranked PAVA and dense/profiled immediate survivor commits retain stage, trace, rollback, and coverage contracts. A previous full-solver prototype changed placement and regressed paired timings, so it was discarded. A next convergence needs those contracts together and exact corpus parity plus performance gates. Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 suite-debt and the no-large-default-closure exit checks remain open.
-
Current-stack CI exposed two new direct imports of private annotation helpers in the n03 test slice. A reviewed public
PlacementContextregression now covers the pattern-only reserved balloon collision as well as the existing dense-table path; the private-import guard, 125 focused tests, static gate, and changed-line coverage (95%, threshold 90%) passed. The 18 PR heads were republished in stack #2088 with refactor source byte-identical to the parity-checked tip. Required CI is restarting on those heads; no stack PR is counted as merged yet. -
Fresh-reviewed PR #2093 extracts Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 hole-queue obstacle evidence and lowers
_place_queuefrom 237 to 182 AST lines. Static/focused checks and 17 exact CAD/STEP drawing/lint/decision/trace snapshots passed; CI is pending. The six named functions and closure exit checks for Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 remain open as a set. -
Fresh-reviewed PR #2091 consolidates three exact lint point formatters (Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932). PR #2092 shares the The same feature is dimensioned twice on one sheet #1599 requirement-dimension build and lowers the issue-named module cap 150→149; PR #2094 removes one genuine two-member clone pair, 53→52 (Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935). Their local static/focused/mutation checks passed and CI is pending. These separate PRs will merge after the dependent stack so main does not invalidate its current CI run.
-
A full local fast-tier preflight of the integrated tip found three failures among 9,904 passing tests: a stale left-strip monkeypatch after the axial owner move, two newly executed detector modules absent from an exact inspection pin, and an unchanged-main lazy evaluator import guard baseline failure. The first two were fixed at their introduction stages in stack #2088 and independently reviewed with load-bearing guards; its refactor source remains byte-identical to the parity-checked tip. Stack #2088 and dependent PRs docs: halve historical issue references in source comments (#1934) #2087/Make rank-zero modules strict package leaves #2089/Move dimension ink producer metadata into typed records #2090/Extract hole queue obstacle evidence phase (#1930) #2093 have been republished on corrected heads and required CI has restarted. A second full local preflight is running; no new merged completion is claimed.
-
Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930's reviewed typed height-rung callback slice is PR #2096, open against the stack; three large default captures (11/6/5) became bound typed state with exact CAD/STEP parity and a mutation-checked guard. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932's reviewed exact-schema-key reuse is PR #2095. The unchanged-main evaluator guard repair is separate PR #2097, with recursive module/class import checks reviewed and mutation-tested. PRs Share exact three-coordinate lint point formatting #2091 and Share expensive requirement-dimension test fixture (#1935) #2092 have passed required CI; they remain unmerged until the dependent stack lands, preventing a main advance that would reset its checks.
-
Latest local gate (2026-09-29): the third full fast-tier run on the current stack tip plus separate Repair lazy STEP evaluator import guard baseline #2097 baseline guard repair passed 9,906 tests, 9 skipped and 15 expected failures in 8m14s (
uv run pytest -q -n 12 --dist worksteal --disable-warnings). The one prior detail-log failure was reproduced as CLI-before-detail logger state; a reviewed test-only fix now targets the actual_corelogger at n06, with its named assertion mutation-tested. Stack #2088 and its dependent PRs were republished once more; all current-head required GitHub CI checks are pending. Share exact three-coordinate lint point formatting #2091, Share expensive requirement-dimension test fixture (#1935) #2092, Consolidate duplicated recognition-refusal test cases (#1935) #2094, and Reuse exact schema key validation (#1932) #2095 have passed their required CI but are held for merge after the stack. The rootmainremains63cb9fe5, and no open PR is counted as merged completion. -
Merged aggregate #2102 at
502bbc46on 2026-09-29, after current-headci-ok, local 9,909-pass fast tier, exact 14 CAD + three STEP drawing/lint/decision/trace parity, and fresh independent review. It integrated 31 reviewed slices through one PR gate; the 31 superseded slice PRs are closed. The rootmainis pulled. Post-merge slow/main workflows are running; no next PR CI has started. -
Merged source census: no module over 3,000 lines and no function over 300. The source-size guard and further typed callback cleanup are reviewed and tested locally for a later single-gate batch; Code health: bring the remaining 33 long functions under 300 lines #1956 remains open until that guard merges. Code health: rationalise rank-0 module sprawl and make module docs track _LAYERS #1933 closed after strict rank-0 and architecture-doc guards passed (22 tests). Code health: strip history/review-round citations from src comments #1934 closed after tokenizer-counted source-comment issue references fell 999→467 and review-round citations reached zero.
-
Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943's reviewed build123d patch passed 2,765 upstream tests and four-workload draftwright parity, but the maintainer chose to keep it local; no upstream PR or CI was started. Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 remains draft in perf: reuse physical lint within finished build policy (#1945) #2085. The remaining Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930/Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931/Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 and performance acceptance work stays open.
-
Merged aggregate #2103 at
12772511on 2026-09-29 after all required current-head checks, includingci-ok, ten Linux shards, Windows and macOS canaries, real-part canary, and coverage. The rootmainis pulled; the prior Integrate reviewed code-health slices with one CI gate (#1939) #2102 post-merge CI completed successfully. A fresh merged-main AST census and four source-shape guards pass: largest source module 2,990 lines (annotations/holes.py), largest function 294 lines (render_balloons), and the six Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 functions 159/168/101/99/178/121 lines. Code health: extract a placement-candidate type and decompose the >1,000-line functions #1930 and Code health: bring the remaining 33 long functions under 300 lines #1956 closed. The issue-named test-module cap is now 148. No ADR text changed. -
Main post-merge CI for Converge placement callback state and guard source shape (#1939) #2103 is queued/running. The next reviewed refactor slices are being integrated locally for one later PR gate; no next PR CI has started. Code health: split annotations/from_model.py (10k lines) per feature family #1929 still requires every annotations module at about 2,500 lines, and Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931/Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937/Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943/Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 remain open. Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 was reopened because Reuse placement critique after fit-advisory removal (#1945) #2048 only covered a partial lint-reuse slice. Its new counted-state candidate remains local pending aggregate review and CI. The build123d Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 patch remains local by maintainer choice.
-
Merged aggregate #2104 at
147f64econ 2026-09-30 after current-head required CI (17 successful jobs, two skipped), local 9,924-pass fast tier, 96.02% total coverage and 99% diff coverage, exact 14 CAD + three STEP drawing/lint/decision/trace parity, and fresh independent review. Code health: split annotations/from_model.py (10k lines) per feature family #1929 auto-closed. The rootmainhas not yet been pulled because shell GitHub DNS is unavailable; the GitHub connector verified the merge. No next PR CI has started. -
The next Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 slices are reviewed and integrated locally. Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 counted-state reuse and Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 typed dimension-spec work remain local while independent review findings and focused regressions are resolved. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937's next seam investigation found distinct PAVA, trace, and transaction contracts; full convergence remains open. Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943's upstream patch remains local by maintainer choice.
-
A smaller follow-up aggregate is ready locally at
9c574776in/private/tmp/dw1939-export-guardafter removing the growing Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 cache experiment. It advances Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931/Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 without claiming their completion. Fresh integrated and PDF-guard reviews cleared; exact 17 CAD/STEP parity,scripts/pr-check --static, a 9,932-pass local fast tier, and quiet paired five-workload no-regression checks passed. The PDF export guard failed under a deliberate registry-resolver omission (180° versus 0°). Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 remains open: its discarded candidate measured only structural traversals, not literal lint calls, and fresh reviews found two state-handling defects. One aggregate PR CI is still the publication plan; no new PR/CI has started. This session's automatic approval policy rejected the draftwright branch creation request, and shell GitHub DNS remains unavailable, so publication is pending restored access. -
2026-09-30 exact-head coverage on the unpublished
9c574776aggregate passed: 9,932 tests, 96.03% total coverage, 98% changed-line coverage (132/134 lines). A later Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 one-build fixture reuse is locally integrated and focused-tested, but has no full-tier timing claim. A proposed Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 four-helper alias was rejected after fresh review found keyword-call incompatibility for a three-line source saving. A direct Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 audit reproduced two lint calls on one returned Box state and three structural lint calls on one unchanged declared Sheet state; literal acceptance is still unmet. GitHub branch creation was rejected again by the automatic approval policy (requires approval, policynever); no new PR or CI started. -
2026-09-30 next unpublished integrated candidate
0f9534c9(/private/tmp/dw1939-after-next-local, based on Refactor annotation owners and share build-local evidence (#1939) #2104): Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 now has zero_dw_source matches with a reintroduction guard; Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 reuses the shared outcome index; Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 lowers the issue-named module cap 147→146; Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 shares the selected leader commit; Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 removes a no-op no-table Sheet hook but remains short of literal once-per-state acceptance. Fresh independent integrated and final-delta reviews found no blocking issue.scripts/pr-check --staticexited 0; the exact-head fast tier passed 9,938 tests, 9 skipped, 15 xfailed, 96.03% total coverage, and 98% changed-line coverage (205/208). The final 14 CAD + three STEP drawing/lint/decision/trace corpus is byte-identical to the Refactor annotation owners and share build-local evidence (#1939) #2104 baseline. Quiet paired five-workload timings showed no reproducible regression; compare-layout Box was 1.001 median next/base over eight fresh pairs after a noisy two-pair 1.010 result, with item/view counts unchanged. No module exceeds 3,000 lines, annotation module 2,500, or function 300. This remains local: automatic approval review rejects GitHub branch/PR writes under policynever; no new PR or CI started, and no issue is counted merged. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937/Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943/Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 and the epic remain open. Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943's upstream patch remains local by maintainer choice. -
2026-09-30 next local dependent follow-up
f8be6b91after frozen0f9534c9: reviewed Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 helper reuse removes 24 net source lines and Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 blind-slot behavior-test consolidation removes 499 test lines while preserving 32 invocations; diagnostic clone pairs 50→40, issue-named cap 146→144. Fresh integrated review found and resolved deleted-file references and a sorted-evidence contract failure. Exact-head static, 9,938-pass full fast tier (96.03% coverage), 100% changed-line coverage, 17-case drawing/lint/decision/trace parity, and paired five-workload no-regression checks pass. One full-tier wall-time run was 440.38 s versus the frozen parent's earlier 435.16 s, so Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's measurable suite-speed acceptance is not claimed. A one-build shared mutable fixture experiment was discarded after review as a poor complexity trade. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937's separate local test-only pre-convergence guardb5badaf9is reviewed and mutation-tested; it will accompany a real engine migration rather than trigger its own CI. Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943's remaining local bbox shortcuts lacked a safe material payoff, and its upstream patch stays local by maintainer choice. Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 literal lint-once remains open because public mutations make persistent caching unsafe; its design audit and scope question are recorded. No new PR/CI or merge: automatic approval review still rejects GitHub branch creation under policynever, and shell GitHub API access is unavailable. One aggregate PR at a time remains the plan. -
2026-09-30 third dependent local unit
636a34d4in/private/tmp/dw1939-third-localfollows frozenf8be6b91: Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 shares identical chamfer/fillet edge evidence (-110 net source lines); Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 folds two blind-slot semantics modules (-406 net test lines, 90 invocations retained), lowers diagnostic clone pairs 40→31 and issue-named cap 144→142, and refreshes 90 CI shard duration keys. A reviewer-found family-specific diagnostic regression was fixed and cleared. Exact-head static, 9,938-pass fast tier (96.02% total coverage), 93.7% changed-line coverage, and byte-identical 14 CAD plus three STEP drawing/lint/decision/trace parity passed; final aggregate review found no P0–P3 issue. One full-tier wall run of 434.49 s versus the parent's earlier 440.38 s is inconclusive for Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's speed acceptance. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937's preselected-commit prototype was discarded after exact parity because an eight-callback adapter saved only ten net production lines and enlarged the hole queue; its reviewed guard remains local. No new PR/CI or merge under the continuing GitHub branch-write approval block. The three local units remain ordered for one aggregate PR CI at a time when publication is available. -
2026-09-30 Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 migration-map and Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 time audit: reviewed docs-only local
1d19f9c8accurately records sparse late, front/rear strip, and dense/profiled immediate paths with the remaining convergence contracts; ADR 2 text and code are unchanged, static/docs gate passed, and Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 stays open. An independent reviewer corrected two initial overstatements before clearing the final diff. A Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 read-only audit rejected a lazy mirror-corpus fixture: targeted 35-test xdist scatters cases, but full collection and CI loadscope normally keep the class together, so the focused speedup would not predict full-tier wall savings. The.test_durationsfile has 1,446 absent-file keys and can be refreshed for CI balance in a separately measured slice. No new PR/CI or merge; local units remain ordered under the branch-write approval block. -
2026-09-30 fourth dependent local unit
fb1a74d9in/private/tmp/dw1939-fourth-localcombines a Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935.test_durationsrefresh and the corrected Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 architecture note. A fresh 12-worker fast-tier run passed 9,938/9 skipped/15 xfailed and measured all 9,962 current cases; previous data weighted 5,013. The 2-way shard estimate on those macOS measurements changed from 2,452.94/2,023.31 to 2,239.59/2,236.66 worker-seconds; this is not Linux CI wall-time proof, so Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's speed acceptance stays open. The docs describe current leader paths and leave Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 convergence open. Combined-head static and independent review passed; no production or test source changed in this unit. The first three unpublished units retain exact 17-case drawing/lint/decision/trace parity, fast tier, coverage and review evidence. Four dependent units are ready locally, but none is a PR or merged; GitHub branch creation remains rejected by the current approval policy, and shell GitHub DNS is unavailable. Publication remains one aggregate PR CI at a time. Remotemainremains merged Refactor annotation owners and share build-local evidence (#1939) #2104 (147f64ec). Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937/Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943/Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 and the epic remain open; Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 upstream patch stays local by maintainer choice. -
2026-09-30 single-gate aggregate publication bundle: All four unpublished dependent units are packaged as one 87-file patch from local
eba73503tofb1a74d9(/private/tmp/dw1939-one-gate-aggregate.patch; PR body and proof manifest alongside it). GitHub Refactor annotation owners and share build-local evidence (#1939) #2104 changed 46 paths from Converge placement callback state and guard source shape (#1939) #2103; local base changed the same paths/rename and all 46 remote merged blob SHAs match, establishing a byte-equivalent base. The patch applied in a fresh base checkout and reproduced the exact aggregate tree9e795307caa1fa86b6b69e5aa520ebb5729c805b. A fresh independent whole-diff review found no P0–P3 issue; 271 focused tests passed; combined changed-line coverage passed at 97% (280/288). Quiet paired five-workload whole-aggregate medians versus the merged-equivalent base were 1.0025, 0.9969, 1.0029, 1.0071, 1.0024 next/base; an eight-pair compare-layout check was 1.0043 median with 0.9772–1.0203 range and no reproducible material regression. These do not prove Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's full-tier speed acceptance or Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943/Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 performance completion. Remotemainremains Refactor annotation owners and share build-local evidence (#1939) #2104. Shellgh apistill cannot connect andgithub_create_branchagain returnedMCP tool call requires approval, but approval policy is never; no PR/CI/merge has occurred. Publication remains one aggregate PR CI gate, with no routing around the approval rejection. -
2026-09-30 single-gate bundle revision 2: Reviewed fifth local unit
99c59d07removes a reproduced-invalid ADR 2 strict xfail, unused dense fixture and stale prose; all 9,961 selected tests retain duration weights. Exact-head static and full fast tier passed 9,938/9 skipped/14 xfailed, 96.03% total coverage; combined changed-line coverage passed 97% (280/288), and a fresh final aggregate review found no P0–P3 issue after 82 focused arrangement/source-size/import/ADR tests. The v2 patch/private/tmp/dw1939-one-gate-aggregate-v2.patch(SHA-256caa596f8ff9cb9af43bbb268639eabcc812404ad1b69fbc784155c19e0b5bd6c) applies to the byte-equivalent Refactor annotation owners and share build-local evidence (#1939) #2104 local base and reproduces Git tree24402bc241cfcbdd696bea26ea3599ebe2b6a243; use/private/tmp/dw1939-one-gate-manifest-v2.mdand its PR body, not the superseded files. A fifth-unit unpaired full-tier wall run was slower (467.87 s versus earlier 434.49 s), so Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 speed acceptance remains open. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932's literal single-definition criterion remains unmet where same-named helpers have distinct semantics; a maintainer scope choice is pending, with no issue wording changed. Shell GitHub API remains unreachable, andgithub_create_branchagain returnedMCP tool call requires approval, but approval policy is never; no PR/CI/merge. Remotemainremains Refactor annotation owners and share build-local evidence (#1939) #2104 (147f64ec), and the one-aggregate-PR CI plan remains intact. -
2026-09-30 whole-epic size check: A fresh
git show/AST census of Python files undersrc/draftwrightcompares the epic baseline71b8c945with merged-equivalent Refactor annotation owners and share build-local evidence (#1939) #2104 baseeba73503: 123→188 modules, 118,103→125,538 physical lines, 10→0 modules over 3,000 lines, and 40→0 functions over 300 lines. The unpublished five-unit bundle99c59d07is 188 modules/125,624 lines with the two large-size counts still zero; its source is net +86 lines versus Refactor annotation owners and share build-local evidence (#1939) #2104. Test Python is 417 files/156,109 lines at baseline, 425/161,464 on Refactor annotation owners and share build-local evidence (#1939) #2104, and 422/161,054 in the local bundle. This establishes better local boundaries but not an overall code-size reduction; module count and total source/test lines remain above the epic baseline. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 fresh read-only review found no viable small deletion: immediate PAVA already usescommit_feature_leader, and direct late-solver migration changes batch spacing/drop, naming, furniture, or trace contracts. Do not add a typed wrapper solely for an apparent convergence. Publish the reviewed bundle through one PR CI gate when GitHub writes are available, then prioritize remaining work by measured deletion, ownership simplification, and reproducible performance evidence. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932/Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935/Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937/Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943/Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 and the epic remain open; Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 local acceptance is still unmerged. -
2026-09-30 single-gate revision 3: A sixth, test-only Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 unit
a2588358reuses the existing read-only CTC-01 annotated fixture in one declared PMI source-identity test and refreshes its duration weight. A direct two-option build found the same nonempty 31 source IDs; the warm targeted test call fell 38.18→5.04 s. Exact candidate fast tier passed 9,938/9 skipped/14 xfailed with 96.03% coverage and static passed; all 9,961 collected fast cases retain weights. Fresh independent review cleared the delta. The combined patch/private/tmp/dw1939-one-gate-aggregate-v3.patch(SHA-256b29970e50132ba953d7a6ea9ad2c0a388e9335814a5fd3bda8bb87372426a3ec) replays on the byte-equivalent Refactor annotation owners and share build-local evidence (#1939) #2104 base to Git treecb7ec2afe500340fc95a9657ab7a1f0b23709961; use/private/tmp/dw1939-one-gate-manifest-v3.mdand the v3 PR body. Reversed-order local 12-worker full-tier timing showed warm Refactor annotation owners and share build-local evidence (#1939) #2104 at 394.82 s and the five-unit bundle at 393.85 s, so the earlier cold 454.16→393.85 s comparison was misleading and Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 full-tier speed acceptance stays open; the sixth unit's single full run was 385.52 s, not a reproducible proof. Remotemainremains Refactor annotation owners and share build-local evidence (#1939) #2104; GitHub branch creation remains rejected by automatic approval policy (requires approval, policynever) and shell API access is unavailable. No PR/CI/merge; one aggregate PR CI at a time remains the publication plan. -
2026-09-30 acceptance-boundary check: Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's fresh green full-tier duration report identified duplicate five-part annotation-box builds, but an independently reviewed class-fixture prototype did not improve paired 12-worker full-tier wall time:
workstealparent/candidate 368.77/374.63 s and CI-styleloadscope416.67/418.96 s. The one-test fold had lost independent failure reporting; the corrected fixture added mutable shared model state for a targeted but noncritical build saving. It stays isolated and is excluded from revision 3. Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935's measured full-tier acceptance remains open. A fresh Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 call-graph review found no small correct route to literal lint-once while public Drawing fields and annotation objects remain directly mutable, placement needs structural critique, and declared physical recognition must stay lazy; the narrower scope question remains pending. No new PR/CI/merge or ADR text. The reviewed revision 3 one-gate patch and publication block are unchanged. -
2026-09-30 Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 critical-path correction: A green uninstrumented 9,938-pass, 12-worker
workstealrun logged all 9,961 selected test completions. The 12 workers finished within 5.1 s (363.1–368.2 s), so a ~30-worker-second one-test fixture saving has only ~2.5 s ideal wall impact and cannot credibly establish the epic's measurable full-tier speed requirement. This explains the discarded annotation-box experiment; prioritize a larger measured aggregate reduction rather than adding isolated shared mutable fixtures. Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 remains open. The revision 3 publication bundle is unchanged, and GitHub publication is still blocked by the branch-write approval policy and shell API failure. -
2026-09-30 reviewed single-gate bundle revision 4 and Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 acceptance evidence: Local seventh unit
7552b9d2reuses the existing lint coverage-state owner across seven families, deleting 65 net production lines; the combined unpublished bundle is now net +21 production lines versus byte-equivalent merged Refactor annotation owners and share build-local evidence (#1939) #2104. Exact-head static, 273 focused tests, 9,938-pass fast tier with 96.02% coverage, 97% aggregate changed-line coverage (295/303), two load-bearing blind-slot mutation failures, and byte-identical 14 CAD plus three STEP drawing/lint/decision/trace signatures passed. Fresh independent slice and whole-bundle reviews found no P0–P3 finding. The v4 patch/private/tmp/dw1939-one-gate-aggregate-v4.patchreplays to Git tree85d935c87de211f12c58a2306ea40300072933fe; its manifest and PR body are alongside it. For Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935, a green original-epic-baseline subset excluding exactly 29 historical failure/error nodes passed 9,564 cases in 500.42 s with 95.69% coverage; the exact current full tier passed 9,938 in 403.13 s with 96.02% coverage under the same 12-worker covered command. This supports all three Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 written measures locally (issue modules 151→142, time lower, coverage higher), with the baseline exclusions and non-attribution caveat recorded in Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935; keep the issue open until publication/current-head CI. Remote main remains Refactor annotation owners and share build-local evidence (#1939) #2104. A freshgithub_create_branchretry still returnedMCP tool call requires approval, but approval policy is never; shellgh apicannot connect. No new PR, CI or merge. One aggregate PR CI gate remains the publication plan; do not route around the approval rejection. -
2026-09-30 merged single-gate bundle and maintainer decisions: PR Consolidate typed evidence and suite debt (#1939) #2105 merged at
f4aa3214after 17 successful current-head CI checks and two expected skips. The mergedmainhas zero_dw_source matches; Code health: replace _dw_* setattr side-channel on helper objects with typed state #1931 is closed. The issue-named test count is 142; Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 retains the two-member clone-budget gap despite the local wall-time and coverage evidence. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 now requires one implementation per equivalent behavior, not one per same name. Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 now targets repeated lint on unchanged internal state within one build attempt and critique mode; public edits still require fresh lint. The local build123d text-alignment cache candidate for Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 is retired by maintainer decision. No ADR text changed. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932, Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935, Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937, Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943, Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 and this epic remain open. Continue with one PR CI gate at a time. -
2026-09-30 Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 scope decision: The maintainer deferred hole/feature leader convergence until after this epic. This goal continues with refactors that preserve successful drawing output. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 remains open as a separate follow-up; the documented current split is not an ADR 2 exception or a completed convergence. No ADR text changed.
-
2026-09-30 merged cleanup gate: Consolidate equivalent lint helpers and evidence tests (#1932, #1935) #2107 merged at
b5ba5564after 17 successful exact-head CI checks and two expected skips. It removes 23 net production and 142 net test lines, preserves test IDs, and adds same-fixture preconditions to the consolidated mutation tests. Independent reviews cleared the final head. Code health: consolidate duplicated contract/lint helpers (finite-real, vectors, _rounded, _index_evidence) #1932 and Code health: shrink test_issue_* modules, close the clone-budget pair gap, share build fixtures #1935 are closed with acceptance evidence and the historical baseline caveat on their issues. The remaining in-goal open work is Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943, Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 and deferred Code health: audit and remove proven dead code paths #2106; Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 is a separate post-epic follow-up. No ADR text changed. -
2026-09-30 Code health: audit and remove proven dead code paths #2106 audit gate: Vulture 2.16 at 60% confidence yielded four proven unused private helpers; independent reviews, 258 targeted tests, static checks and byte-identical fresh 23-case drawing signatures passed. PR Remove proven dead private drawing helpers #2108 is the sole active epic CI gate. The raw scanner report remains advisory and false positives are recorded on Code health: audit and remove proven dead code paths #2106. Old draft perf: reuse physical lint within finished build policy (#1945) #2085 was closed after fresh Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 counted-call and timing evidence found no useful complete patch; Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 and Perf: lint each drawing state at most once (12 lint passes on grm03) #1945 remain open with measured acceptance unmet.
-
2026-09-30 Code health: audit and remove proven dead code paths #2106 merged: Remove proven dead private drawing helpers #2108 merged as
a243f004with 17 successful exact-head CI checks and two expected skips. Code health: audit and remove proven dead code paths #2106 is closed. The Vulture audit is advisory, its false positives are documented, and no blanket gate was added. Remaining in-goal issues are Perf: cut optimal bounding-box and deepcopy cost in annotation/text geometry (~25-33% of samples) #1943 and Perf: lint each drawing state at most once (12 lint passes on grm03) #1945; both retain unmet measured performance acceptance after fresh current-main investigation. Code health: decide whether hole and feature leader placement should share one engine (ADR 2) #1937 remains deferred outside this epic.
2026-09-30 closure decision
The maintainer chose to close this epic now. Its completed structural, smell, documentation, test and accepted performance slices are merged and recorded above. #1943 and #1945 are measured exceptions, detached from the epic and left open as separate performance follow-ups. Neither issue's original measured acceptance is claimed; the current-main profiling, rejected experiments and remaining technical constraints are logged on those issues. #1937 was separately deferred after the maintainer chose to keep successful drawing output unchanged. No ADR text changed for this closure, and no new optimization or CI run was introduced.
Activity
17 remaining items
pzfreo commentedon Sep 29, 2026
Runner-efficient integration update (2026-09-29): the 27 active per-PR Actions runs for this epic were canceled after the maintainer reported that the full stack matrix was taking shared runners from other projects. No other project runs were touched.
mainrequires strict current-headci-okfor each PR, andgh stack mergecannot reuse the tip result for lower PRs. Therefore reviewed slices are now collected in one integration PR, #2102, againstmain; the original small PRs remain review records. Draft performance #2085 is excluded.Before opening #2102, the integrated head
feb38bc4passed static checks and the full local fast tier (9,909 passed, 9 skipped, 15 xfailed in 7m07s). Exact 14 CAD + 3 STEP drawing/lint/summary/decision/trace JSON matched the prior refactor-stack tip byte for byte. A fresh independent integration review found no P0–P3 findings and reran 101 targeted checks. This PR now has the only active epic GitHub CI matrix. It will squash-merge after its required current-headci-okpasses; the source PRs will then be closed as superseded. No open work is counted as merged yet.Integrate reviewed code-health slices with one CI gate (#1939) (#2102)
pzfreo commentedon Sep 29, 2026
Status (2026-09-29): aggregate #2102 merged at
502bbc46; its post-mergemainCI run 36627991793 completed successfully. To keep PR CI to one head, the next 21 reviewed commits are in PR #2103, based on502bbc46, with auto-merge configured only after required checks pass. Localscripts/pr-check --fullone5b390e5passed 9,913 tests, 96.01% total coverage and 98% changed-line coverage; 14 CAD + 3 STEP drawing/lint/decision/trace records match merged main exactly. A Windows-only test encoding correction atbb8e28ddpassed focused/static checks and fresh review; required current-head CI is running. It adds guards for the epic's 3,000-line module / 300-line function / five-default closure exits and the six #1930 <200 functions, folds one issue-named test module (149→148), and extracts polygonal prism jobs. #1930 and #1956 close on merge if CI passes. No ADR text changed.#1929 was reopened because its own ~2,500-line annotation-module acceptance is still unmet on
main(four modules over that threshold). Separate #1931, #1932, #1937 and #1929 follow-on slices remain local and unmerged. The build123d #1943 patch remains local at the maintainer's request; #1945 remains draft pending its no-regression gate.Read source-shape inputs as UTF-8 on Windows (#1939)
Converge placement callback state and guard source shape (#1939) (#2103)
pzfreo commentedon Sep 29, 2026
2026-09-30 status after merged #2103 (
12772511): its required PR CI passed, and post-mergemainCI has passed lint, view-selection, remaining-2, and CTC-04 slow lanes; onlyremaining-1is still running with no failure reported. The priormainrun took about 90 minutes on that lane, so no next PR CI has been started while it occupies a runner.The next single-gate aggregate is ready locally only at
eba73503against merged main: fresh integrated architecture review clear; independent lint review found no other source defect and confirmed #1945 remains partial.scripts/pr-check --fullexited 0 (9,924 passed, 9 skipped, 15 xfailed; 96.02% total and 99% changed-line coverage). Exact 14 CAD + three STEPbuild_drawingdrawing/lint/summary/decision/trace comparisons match main 17/17. All 39 annotation modules are ≤2,500 lines and the new size guard failed under a deliberate 2,507-line mutation. Alternating paired builds across GRM03, blind-hole STEP, explicit/compare Box, and holed-slot found no reproducible combined regression. This aggregate will close #1929 after its own current-head CI passes; #1931/#1932/#1935/#1937/#1945 remain open. The literal #1945 double-lint gap and sequenced follow-up are recorded on that issue.A separate later local batch already combines reviewed #1931 span, #1937 shared leader commit, and #1932 exact-index slices with static/focused checks and 17/17 parity; it is not published or counted as merged. The maintainer's build123d #1943 patch remains local and unpublished as requested. No ADR text changed.
Refactor annotation owners and share build-local evidence (#1939) (#2104
Consolidate typed evidence and suite debt (#1939) (#2105)
pzfreo commentedon Sep 30, 2026
2026-09-30 status: structural/code-health work in this epic is merged, including #2108 / #2106. The two remaining in-goal issues, #1943 and #1945, do not meet their original measured performance acceptance. Fresh current-main four-workload profiling and independent audits found no material, output-preserving bbox/deepcopy patch in Draftwright-owned paths; the retired upstream text-cache candidate and smaller cache experiments did not justify a change. Fresh counted-call review found that #1945's opaque post-build hook can return the same object whether it changed lint-relevant state or not; a complete lint-once guarantee needs mutation tracking or a hook contract, and the small table-only patch had no reliable useful timing gain. No additional patch or CI run was started.
The maintainer has been asked whether to close #1939 with #1943 and #1945 explicitly left open as measured exceptions outside the epic, or keep the epic open until their targets are met. Until that decision, the epic stays open and neither performance acceptance is claimed.
pzfreo commentedon Sep 30, 2026
Closed by maintainer decision after all 17 retained subissues completed. #1943 and #1945 remain open as separate measured performance follow-ups; their original acceptance is explicitly unmet, as logged in this issue and each follow-up. #1937 remains a separate deferred follow-up. No ADR or successful drawing behavior changed for this disposition.