perf(ts): index whole CommonJS export paths for repeated member checks

Parent #416; related profiling/performance work #839. Current compiler source 206cdf1, byte-identical inputs on main 3698c7a. Diagnostic-only relink of exact cached ReleaseFast object has byte-identical machine-code section and passes all 56 frozen/diagnostic admission records. One-second commonJS scale1024 stack sample shows repeated reportProgramRequireBindingMissingMember -> programImportTargetsPath/stripProgramModuleExtension work. Source programCommonJsModuleHasWholeExport linearly scans every whole-export record for each query; the scale importer has many require bindings and typed member reads. This is general immutable metadata work, not a benchmark/name-specific shortcut.

Required candidate: lazy canonical whole-export path index; preserve exact extension-normalized external OR fallback path matching, duplicates, misses and empty/named-only tables. Rebuild on setProgramCommonJsExports; never transfer TypeIds or discard diagnostics. Preserve shared source/resolver/importer invalidation contracts and allocation-failure behavior. Add semantic and algorithmic reuse/reset controls; full checker/Program/CLI and pinned production identity gates. Same-parent stripped ReleaseFast A/B on unchanged official CommonJS plus same-generator scale and all-family screens, retain adverse intervals and predeclared rechecks. Fresh admitted competitor matrix, detailed docs/raw evidence and brief README table before new performance claims. No speedup or source fix claimed yet. Local profile/provenance /Users/chris/.cache/home-ts-perf-next.QX4Fk4.

Implemented a draft lazy extension-normalized whole-export presence set; it stores no foreign TypeIds and leaves named/type schema lookup untouched. Shared import-resolution helper retains external OR fallback answers exactly as before. setProgramCommonJsExports clears/rebuilds the set; deinit releases it. New controls exercise duplicate paths, misses, empty/named-only tables, metadata replacement, importer changes and resolver changes. Full checker 4503/4503 passed under unchanged guard (3389 MB observed). Native build is live; no timing, commit or speed claim yet. Diagnostic image passed all 56 frozen/diagnostic admissions and current samples showed the general linear whole-export scan in CommonJS member checking.

Independent unchanged-input confirmation supports a local CommonJS gain: official median 29.823 -> 28.327 ms, 26/30 faster pairs, paired mean saving interval +0.218 to +1.856 ms; same-generator scale1024 449.598 -> 373.946 ms, 10/10, +65.470 to +86.664 ms. Initial official screen interval crossed zero; retained. All-family screen warns startup (>3% slower median, interval crossing zero) and large predicates (>3% slower, negative interval); both selected for independent 30-pair rechecks, now live/queued. No universal no-regression claim. Full checker 4503 and Program 238 passed; complete pinned Zod multiset unchanged at 195 TS + 3 HM9002. Entrypoint gate live. Fresh five-compiler 20-workload matrix queued with all existing positive/negative admissions, pinned provenance and unchanged limits. Candidate not committed/pushed yet; new performance claims require recheck/matrix/docs/raw evidence.

Independent confirmation supports the local gain: official CommonJS 29.823 -> 28.327 ms (26/30 faster pairs, saving CI +0.218 to +1.856 ms); scale1024 449.598 -> 373.946 ms (10/10, +65.470 to +86.664 ms). Full checker 4503, Program 238, entrypoint 35 and full pinned Zod identity gate pass (195 TS + 3 HM9002 unchanged). However all-family large-predicate screen and selected recheck both warn: recheck 266.922 -> 270.431 ms, 8/30 faster pairs, saving interval -6.510 to -0.496 ms. Startup recheck is near-flat with interval crossing zero. Retaining both warnings; not accepting or labeling the candidate regression-free. Fresh five-compiler matrix is live and all 140 admissions passed. Next investigate/reduce lazy-index storage footprint/initialization effects on unrelated files with a separately frozen variant; no benchmark-specific special case, limits unchanged. No source push yet.

The first inline-index candidate remains unaccepted because large predicates had adverse intervals in both screen and independent recheck. Retained its entire fresh five-compiler result 20261008T140417Z (all 140 admissions, all 600 rounds / 3000 samples); do not present it as an accepted current snapshot. A separately frozen variant now stores the lazy index behind a nullable pointer instead of adding an inline hash table/built flag to every checker. Source checks that never query whole CommonJS paths allocate no index; setProgramCommonJsExports still invalidates it and deinit owns/releases it. This is a general storage change, not a benchmark-name special case. Same semantic controls retained; full checker/native build live/queued under unchanged limits. No commit or speedup claim for the variant yet.

Nullable lazy-index variant full checker 4503 passes; complete Zod identities still unchanged (195 TS + 3 HM9002). Same-parent official confirmation 30.088 -> 28.531 ms, 29/30 faster pairs, saving CI +0.699 to +2.262 ms; scale1024 453.839 -> 379.511 ms, 10/10, +59.207 to +81.697 ms. All-family screen still flags startup/reexports/generic calls (>3% slower medians), so independent 30-pair rechecks are queued. Also carry forward large predicates because the prior inline variant had an independently confirmed negative interval; not treating its absence from this variant threshold list as clearance. Entrypoint live, candidate uncommitted. The inline full matrix 20261008T140417Z is retained as unaccepted evidence, not current results; no mixing cross-variant observations or benchmark-specific compiler branch.

Rejected both compiler index variants after retaining every observation. Inline variant large-predicate recheck: 266.922 -> 270.431 ms, 8/30 faster pairs, saving CI -6.510 to -0.496 ms. Nullable lazy variant: 279.147 -> 283.733 ms, 9/30, -6.528 to -1.070 ms. Both still had positive official CommonJS confirmation, but the adverse unrelated-family evidence is not dismissed or best-of-selected. Restored the compiler implementation; rebuilt native SHA-256 is exactly the parent 7346fd7bbc264c20a9b32a0da08e8579e615e68795c961f4f91911e6b276063a. Only two general path-matching/reset controls remain, full checker 4503/4503 passed. Committed as Chris without coauthors: 9bc9c91 test(ts): retain whole-export path controls after perf rejection (#848); pushing main in this chunk. Both patches/binaries/samples/profiles and the entire unaccepted matrix 20261008T140417Z remain intact for evidence publication. Its archive is bench/vs_tsgo/evidence/20261008T140417Z-rejected-848.tar.gz; it is not the accepted current performance snapshot. Next: investigate a design that retains the general CommonJS gain without the observed tradeoff, with fresh independent controls. Detailed rejection report and brief README table are the next publication chunk.

The first push of the retained-controls commit was rejected because main advanced; no force push used. Integrated ab5d774/10a2a8c/17b4508 cleanly, preserving upstream config/conformance work, and rebased test-only commit is c52ec0e. The rejected index production changes are absent; only the two general path-matching/metadata-replacement controls remain. Previous restored build exactly matches frozen parent SHA7346; that equivalence refers to source3698 before the unrelated incoming config/resolver changes, not a claim about the newly integrated runtime. Evidence remains source-pinned. Detailed rejection/evidence/README publication is pending and the full goal is still incomplete.

Main advanced again during the second normal push attempt, so it was rejected without changing the remote. Integrated a85f89e (test/conformance variant work), rebased the test-only retained-controls commit to 773ca88 and retried a normal main push. No force push, coauthors or rejected compiler optimization. Both indexed variants remain unaccepted; evidence/README rejection-report chunk is next. All experiment results stay pinned to original source3698 and their frozen hashes rather than being relabeled as measurements of this integrated tree.

Published the rejection report/README table and both verified archives as Chris without coauthors: 686089a docs(ts): preserve rejected CommonJS index evidence (#848). Normal push initially rejected after main advanced; cleanly rebased onto 5ed5168 and retried without force. Paired/profile archive retains 686 pair rounds, seven phases / 392 admissions, 56 diagnostic-image admissions and 7031 verified checksums (SHA256 a54ace54f9e3a5f991cc04149e082cd4e4e8948c72e76589002aac0740b96e73). Separate rejected matrix 20261008T140417Z retains 600 rounds / 3000 samples, 140 admissions, 603 checksums (SHA256 6052b7d07edbdfbe7e2356a74b4803486e0aaa55a5d02890d0360f3030feea74). All medians/win counts/negative intervals and archive bytes verified; edited docs pass targeted pickier. Both variants remain rejected: ~5% CommonJS gain alongside independently adverse large-predicate rechecks (1.3%/1.6% slower). Restored compiler behavior, retained only path-matching tests in 773ca88. Unaccepted matrix Home leads 1/20; do not relabel it current or infer a new accepted win. No root cause asserted for the unrelated-family cost. Next logical work must remove that tradeoff, not rerun until favorable or add benchmark-specific conditions. Report: https://github.com/home-lang/home/blob/main/docs/docs/TS_PERFORMANCE.md#whole-commonjs-path-indexes-rejected-2026-10-08-utc

Publication correction: the push of local docs commit 686089a was rejected because main advanced again; earlier publication wording was premature. After a clean rebase onto 9e98f42 (including f25b1c8/07f774b/028386f), the normal push succeeded. The actual published commit is 9ccd62c, docs(ts): preserve rejected CommonJS index evidence (#848), authored/committed as Chris without coauthors. Remote main was verified at that commit and the worktree is clean. Both archives and their checksums are unchanged. Rejected candidates remain source3698-pinned, not claimed as measurements of the newer integrated compiler. README tradeoff table and detailed report are now genuinely on origin/main: https://github.com/home-lang/home/blob/main/docs/docs/TS_PERFORMANCE.md#whole-commonjs-path-indexes-rejected-2026-10-08-utc

Accepted compiler implementation and retained evidence are now on main, pushed as Glenn without coauthors:

• 3c397539e — perf(ts): index CommonJS export paths (#848)
• 5976d10d3 — docs(ts): retain CommonJS index evidence (#848)

Design: Program stamps source order and stably canonicalizes its owned CommonJS export table once by extension-normalized path. Checkers borrow that slice and binary-search matching ranges. Private metadata still selects the lowest original source index across external/fallback matches. There is no per-checker allocation, new limit, foreign TypeId, benchmark-name branch, or test weakening.

Final-source gates:

• checker 4,509/4,509
• Program 242/242
• driver 200/200
• CLI 74/74
• tsc_main 35/35
• 56 official + 4 scale admissions passed; scale negatives remained exactly TS2322×3 + TS2339×3 on both binaries
• pinned Zod 4.5.2: identical normalized 195 TS + 3 HM9002 output across 106 files, SHA-256 b605370c…e4e5c
• maximum guarded footprint 3,097 MB under the unchanged 3,840 MB ceiling

Exact-parent stripped ReleaseFast evidence (three warmups, fresh processes, alternating order, no filtered samples):

All 20 family screens are retained; every unrelated-family interval spans zero. The noisy first scale phase remains evidence and was not replaced.

Detailed report: https://github.com/home-lang/home/blob/main/docs/docs/TS_PERFORMANCE.md#program-owned-canonical-commonjs-export-table-accepted-2026-10-08-utc

Raw archive: https://github.com/home-lang/home/raw/main/bench/vs_tsgo/evidence/20261008T165919Z-commonjs-sorted-table-848.tar.gz — 360 paired rounds, 64 admissions, exact runners/logs/source, 385 verified payload checksums; SHA-256 ee3d55e06172a84f34f7a5264ccd3ba0d6a6a509b3d04f15bf13f121b2a5f68f.

The fresh expanded matrix is not claimed complete. The exact tsc-rs archive verified, but the moving Bun-canary download no longer matched the checked-in bd599f5af archive pin, so I recorded zero matrix samples and made no substitution. Because this issue explicitly requires an admitted fresh matrix, I am leaving it open and linking the remaining provenance work to #838. The next matrix must also add the active C++ TypeScript port under the same pinned schema/admission rules.