ReviewOS

home-lang/home

Track TypeScript frontend benchmark leadership

#416
Open chris opened this 23 days ago · 283 comments
23 days ago

Goal

Maintain a fair, reproducible TypeScript frontend performance suite and improve Home until it leads JavaScript tsc 6 and stable native TypeScript 7 across every accepted workload.

Benchmark rules

  • All compilers receive identical valid source graphs and compiler options.
  • Every compiler must validate successfully and silently before timing, and reject the same invalid controls with the expected diagnostics.
  • All selected workloads must pass admission before any timing or result-directory creation; no failing row is silently omitted.
  • Compiler versions and generated workloads are pinned.
  • JavaScript TypeScript 6 and native TypeScript 7 are installed in isolated prefixes.
  • Measurements rotate compiler order within every round to balance workstation load.
  • Workloads represent general compiler behavior, never benchmark-specific shortcuts.
  • Results, environment details, caveats, and reproduction commands stay documented.

Landed foundation

  • Nineteen deterministic workloads covering startup, files, complex types, module graphs, TSX, inference, control flow, predicates at two scales, null-safe access, destructuring, overloads, classes, structural objects, interface composition, variadic tuples, checked JavaScript, and recursive generic consumers
  • Pinned JavaScript tsc 6.0.3 and stable native TypeScript 7.0.2 toolchain
  • Interleaved Hyperfine measurement schedule and silent-success validity gate
  • Release allocator, reusable source I/O, and bounded large-program parallelism
  • Cached source facts, bind-only export analyses, and projected export-name sets
  • Indexed lexical declarations, interface merge predecessors, namespaces, annotations, and validated type display names
  • Pruned provably irrelevant overload, inference, import, expando, and compatibility work
  • Relation-cache working-set retention and single-probe insertion
  • Full performance documentation and linked README summary

Current correctness and measured status

Implementation 1cfc33c (#536), audit d95dfaa (#541), and verified README/report 707045c are pushed to main as Chris with no coauthors. TS 6.0.3 and native TS 7.0.2 are verified; TS7/tsgo is one competitor.

The prepared-owner CommonJS display query removes the extra full source compilation and raw-text recovery. Program140/140, conformance1418/1418, CLI69/69 and harness75/75 pass; unchanged checker/driver targets verify from cache. All19 unchanged admission gates and all15 complete audit case sets preserve every previously passing case. Disk-interrupted logs remain alongside complete retries; no partial audit was accepted.

Same-correctness 30-round alternating comparison: import graph32.4±1.7ms to28.4±0.6ms, 12.4% less time and30/30 paired wins. Barrel27.9±5.2ms to27.3±4.2ms, only2.3% lower with17/30 paired wins: inconclusive, not a reliable improvement. All60 rounds/120 successful samples retained.

Full measurement20260828T234350Z validates570 rounds/1710 successful finite samples, with all513 input files byte-for-byte unchanged. Home has lower means on19/19: import graph28.1ms vs TS7 47.0ms (1.67x), barrel25.6ms vs42.7ms (1.67x), large predicates340.5±18.2ms vs362.2±30.4ms (1.06x;29/30 paired wins). Several rows have substantial shared-host variance. Independent confirmation20260828T234818Z retains30 rounds/90 samples: large predicates Home358.6±11.7ms vs TS7 388.5±16.7ms (1.08x;30/30 paired wins). It is separate, not substituted or pooled. Selected compiler/native-payload/runtime/launcher paths and hashes match before and after both measurements; automated harness provenance is tracked by#544.

CommonJS instance typing is NOT admitted for timing: TS6/TS7 pass66/66, Home22/66, with all44 missing wrong-type/missing-member rejections retained under#541. Export-list controls remain80/96 (#540). Other remaining gaps include owners12/20, nominal origins44/52, globals32/56 and bound globals44/56. The broader goal stays open: real projects, additional rejection coverage, other platforms and semantic correctness remain unproven. Next priorities are actual source-owned CommonJS require types#541 and precise external hidden-binding diagnostics#540.

Frozen candidate commonjs-query.hyU20d/release-v1/bin/home-tsc, SHA-256573257a819a178c9387f6d02a7992c08f995185f5dbc94034eb5eadc0ae956c0. Full results, retained history and reproduction.

Historical indexed-export checkpoint

Implementation b7b9dd9 (#537) and parser ownership fix d6878e8 (#538), with verified README/report 24a34f4, are pushed to main. Compiler versions are verified TS 6.0.3 and native TS 7.0.2; TS 7/tsgo is one competitor.

All 19/19 unchanged workload admission gates pass. Full measurement 20260828T231439Z validates 570 rounds / 1,710 successful finite samples; fresh generation matches 513 measured input files. Home has lower means on 19/19 workloads, but large predicates have only a noisy 1.02x lead: 371.6 ± 125.5 ms versus TS 7 377.7 ± 66.4 ms, 25/30 paired wins. Independent 30-round confirmation 20260828T231856Z retains all 90 samples: Home 320.4 ± 2.8 ms versus TS 7 331.4 ± 4.6 ms (1.03x; 29/30 paired wins). Confirmation is separate, not substituted into the full table.

The same-correctness alternating before/after comparison for #537 improves 525.6 ± 12.4 ms to 353.5 ± 10.5 ms, 32.7% less time and 30/30 paired wins. Deterministic source visits drop 65,536 to 256 for 128 independent exports. Bound-owner edge indexing preserves semantic query work; no control or failed observation was removed.

New exported-variable-list audit improves 0/96 to 80/96. All 32 positives, 32 wrong-type controls and 16 barrel hidden-name controls pass; 16 direct-hidden diagnostics remain TS2305 instead of TS2459 (#540). Existing audit scores remain: factory 240/240, recursive 288/288, generic classes 120/120, static 84/84, bound classes 52/52, callable identity 56/56, callable unions 256/256, graph discovery 28/28, export origins 32/32, owners 12/20, nominal origins 44/52, globals 32/56, bound globals 44/56. Both competitors pass every audit.

Verified suites: checker 4286, conformance 1417, Program 137, parser 891, driver 187, CLI 69, emitter 499 and benchmark harness 72 tests all pass. Pickier reports no errors and one pre-existing README fragment warning. Frozen evidence export-index.NVcIdu, SHA-256 615e9fa1f78c2cad17a0b8911b897ede73c00422138d6ecbb2b74801d4564f14.

Next: #536 removes redundant CommonJS export-query compilation; #540 adds external owner-local diagnostic facts. Correctness gaps, expanded rejection coverage, real projects and cross-platform measurements keep the overall goal open. Shared-host synthetic results are not universal benchmark leadership. Full results and reproduction.

Historical exported-factory checkpoint

Implementation de8fe28 (#534), documentation/README f3740df. Executables are verified TS 6.0.3 and native TS 7.0.2; native TS7/tsgo is one competitor.

All 19/19 workload admission gates now pass for all three compilers. Both original graph projects and their unchanged TS2322/TS2339 rejection controls pass; no failing workload was removed. The 240 exported-factory controls improve from Home130/240 to240/240. Recursive288/288, generic classes120/120, static84/84, bound classes52/52, callable identity56/56, callable unions256/256, graph discovery28/28 and export origins32/32 remain passing. Imported owners improve10/20 to12/20; nominal origins44/52, globals32/56 and bound globals44/56 remain incomplete. Both competitors pass all these audit cases.

Full relevant suites: checker4286/4286, conformance1417/1417, Program133/133, driver187/187, CLI69/69, source-owned focus10/10, harness69/69.

Complete measurement 20260828T224220Z: 30 interleaved rounds after3 warmups, all570 round files /1710 samples validate; fresh generation matches all513 measured input files. Shared Apple M3 Pro workstation with unrelated CPU/filesystem activity. Home has lower means on 18/19 workloads, not universal leadership. Import graph33.7ms vs TS7 50.8ms (1.51x); barrel graph25.1ms vs41.1ms (1.63x). Large predicates lose: Home541.3 ±50.7ms vs TS7 380.6 ±44.8ms, 1.42x slower and0/30 paired wins. That losing row is retained in README and full docs.

Next: #537 establishes and fixes the large-predicate scaling cause against a same-correctness controlled baseline; #536 removes redundant CommonJS export-query compilation. Remaining semantic failures, broader rejection coverage, real-project and cross-platform measurements keep the overall goal open. Frozen evidence: graph-types.4EsbvV; binary SHA256 9f5e333beaf9828a9720ce260ae05a872ebcea3bb51bf5785ffa3ccd7b72f1bc. Documentation: docs/docs/TS_PERFORMANCE.md#exported-factory-checkpoint-performance.

Historical checkpoint before exported-factory typing

Checkpoint: implementation 3b7f0ea (#524), recursive throughput workload de72563, report-integrity fix 74d1163, and documentation/README 91d6ca8. Pinned executables are TS 6.0.3 and native TS 7.0.2; TS 7 and tsgo are one competitor.

The original admission set remains 16/18 for Home: both imported graph rejection gates fail and were not timed. The added recursive workload passes its positive and nine appended negative controls, making expanded admission 17/19 (both competitors 19/19). No failure is removed by expanding the denominator.

Untimed correctness gateTS 6.0.3Native TS 7.0.2Home
Global declarations56/5656/5632/56
Bound-global discovery56/5656/5644/56
Imported owners20/2020/2010/20
Nominal identity52/5252/5244/52
Imported generic classes120/120120/120120/120
Recursive generic consumers288/288288/288288/288
Import/re-export graph admission2/22/20/2

Lazy source-owned definitions improve recursive controls from 112/288 to 288/288 and imported generic classes from 118/120 to 120/120. Local non-exported classes remain non-exported; transparent aliases preserve their target's origin. No fixed-depth unfolding or new permissive fallback. Inheritance, global/owner typing, broader generic forms and original graph linkage remain open (#521, #480, #487).

Fresh full timing 20260828T221149Z has lower Home means on 17/17 admitted timed workloads, with 510 complete round files and 1,530 successful samples. Apple M3 Pro, 30 interleaved runs after three warmups, shared-workstation load. Recursive generics: Home 29.7 ± 0.2 ms, native TS 7 71.8 ± 1.1 ms, TS 6 151.7 ± 2.2 ms (2.42× lower mean). Large predicates: Home 377.0 ± 100.9 ms, native TS 7 393.6 ± 94.1 ms (1.04× lower mean; noisy). Other rows also have substantial variance.

Independent confirmation 20260828T221611Z retains 60 complete rounds / 180 successful samples. Recursive generics: Home 29.6 ± 0.7 ms vs native 73.1 ± 3.9 ms (2.47×; Home faster in 30/30 paired rounds). Large predicates: Home 349.9 ± 17.0 ms vs native 392.7 ± 83.0 ms (1.12×; 28/30 paired rounds). Confirmation does not replace or get averaged into the full run. The predecessor fails recursive rejection controls and is ineligible for a before/after timing claim.

Exact-source tests: checker 4285/4285, conformance 1417/1417, Program 131/131, driver 187/187, CLI 69/69, focused source-owned 9/9, harness 66/66. Release SHA-256: 267691e33056e2c4a8b23b757bc3eed19c8b3d90fa9407120b8985834a844ccd. Reporting now rejects incomplete/extra rounds, incorrect compiler coverage/order, malformed JSON, and failed/invalid samples instead of publishing partial averages.

Full results, variance, preserved history and reproduction. Correctness details.

Not complete: these are local synthetic observations under shared-host load, not universal, real-project or cross-platform leadership. The two original graph gates and broader semantic coverage still fail or remain unverified. The tracker stays open.

Historical measured snapshot (not current admission)

Apple M3 Pro, 30 interleaved runs after three warmups; raw full-suite result 20260827T202656Z, compiler commit 9e45e10, documentation commit 75e02f3. Every observation is retained, including variance and outliers.

Workloadtsc 6.0.3native TS 7.0.2Home 0.1.0Home vs fastest competitor
startup64.4 ± 1.8 ms42.2 ± 11.3 ms3.9 ± 2.7 ms10.82× faster
many_files213.2 ± 14.8 ms53.5 ± 3.1 ms30.0 ± 1.5 ms1.78× faster
deep_types127.9 ± 2.2 ms52.6 ± 1.5 ms12.9 ± 0.3 ms4.09× faster
import_graph129.2 ± 1.9 ms45.0 ± 0.9 ms25.8 ± 1.1 msIneligible: fails semantic controls
reexport_graph94.7 ± 3.0 ms40.7 ± 0.9 ms26.8 ± 0.6 msIneligible: fails semantic controls
tsx_components159.1 ± 1.2 ms46.3 ± 0.9 ms21.3 ± 0.3 ms2.17× faster
generic_calls177.5 ± 4.4 ms55.4 ± 11.2 ms22.0 ± 1.8 ms2.52× faster
control_flow212.2 ± 127.9 ms63.1 ± 22.3 ms32.9 ± 3.8 ms1.92× faster
type_predicates237.3 ± 14.0 ms72.4 ± 10.8 ms46.5 ± 19.5 ms1.56× faster
type_predicates_large1027.8 ± 117.6 ms348.8 ± 15.7 ms322.1 ± 13.3 ms1.08× faster
null_safe_access206.3 ± 12.4 ms60.3 ± 2.4 ms40.6 ± 1.8 ms1.49× faster
destructuring144.3 ± 11.1 ms48.5 ± 1.1 ms35.3 ± 0.8 ms1.38× faster
overload_resolution222.7 ± 20.6 ms71.6 ± 10.7 ms30.9 ± 2.9 ms2.32× faster
class_hierarchy200.0 ± 17.6 ms55.8 ± 4.5 ms29.8 ± 1.1 ms1.87× faster
structural_objects197.7 ± 9.7 ms60.7 ± 2.8 ms31.5 ± 1.1 ms1.93× faster
interface_composition233.3 ± 43.6 ms76.2 ± 30.3 ms46.5 ± 3.4 ms1.64× faster
variadic_tuples268.4 ± 33.0 ms81.0 ± 8.5 ms46.2 ± 8.3 msProvisional: older admission schema
checkjs_jsdoc246.4 ± 48.7 ms63.0 ± 16.6 ms41.2 ± 6.8 ms1.53× faster

Historical snapshot: 18/18 lower Home means were recorded, with stronger primitive-return checking. This is not current acceptance: the two graph claims were subsequently withdrawn and later compiler changes have not been retimed. Large predicates have a 1.08× mean lead. The predeclared independent 30-round confirmation 20260827T203044Z records Home 342.5 ± 29.3 ms vs native TS 7 370.2 ± 22.0 ms (1.08× lead); TS 6 is 1086.3 ± 96.2 ms. Home is faster than native in 29/30 paired large-case rounds in the full run and 26/30 in the confirmation. The confirmation does not replace or get averaged into the full table. All 1,710 timed samples succeeded; exact versions, balanced ordering, published tables, preserved binary, and frozen predicate hashes were verified.

All 99 program, 4,226 checker, 173 driver, and 16 harness/report tests pass, as do all 18 shared positive projects and 13 negative controls per compiler. Five new return-contract regression groups have been independently checked with the same CLI inputs against both TS baselines (#474). Async is not an accepted workload: #475 and #476 still fail semantic controls, and #478 tracks diagnostic wording. Do not infer full TypeScript parity from these counts.

These are local synthetic results, not proof of universal, real-project, or cross-platform leadership. Several rows have substantial variance; no observations were discarded. The correctness fix is not itself claimed as a performance optimization. Full methods, history, and reproduction.

Previous 18-workload snapshot (retained)

Previous measured snapshot

Apple M3 Pro, 30 interleaved runs after three warmups; raw full-suite result 20260827T195539Z, compiler commit d330198, documentation commit 20084a4. Every observation is retained, including variance and outliers.

Workloadtsc 6.0.3native TS 7.0.2Home 0.1.0Home result
startup66.7 ± 2.5 ms42.0 ± 9.6 ms3.6 ± 0.4 ms11.73× faster
many_files215.4 ± 3.3 ms56.7 ± 7.5 ms32.4 ± 6.3 ms1.75× faster
deep_types146.0 ± 8.6 ms58.3 ± 3.7 ms14.1 ± 0.8 ms4.12× faster
import_graph139.0 ± 4.4 ms49.8 ± 4.1 ms29.3 ± 2.0 ms1.70× faster
reexport_graph111.0 ± 46.4 ms47.0 ± 19.2 ms31.1 ± 6.1 ms1.51× faster
tsx_components178.0 ± 25.2 ms50.2 ± 1.6 ms23.3 ± 4.1 ms2.16× faster
generic_calls188.4 ± 12.9 ms57.3 ± 3.9 ms23.2 ± 2.0 ms2.47× faster
control_flow190.2 ± 4.8 ms59.5 ± 2.2 ms33.6 ± 5.4 ms1.77× faster
type_predicates237.8 ± 12.7 ms70.9 ± 2.0 ms41.9 ± 0.8 ms1.69× faster
type_predicates_large1010.4 ± 25.3 ms351.4 ± 15.2 ms323.8 ± 10.2 ms1.09× faster
null_safe_access191.9 ± 4.1 ms57.2 ± 1.1 ms38.7 ± 0.5 ms1.48× faster
destructuring138.4 ± 1.6 ms47.3 ± 1.0 ms34.4 ± 0.5 ms1.37× faster
overload_resolution203.1 ± 3.4 ms64.6 ± 1.4 ms29.1 ± 0.6 ms2.22× faster
class_hierarchy186.1 ± 7.9 ms53.4 ± 7.8 ms28.1 ± 0.4 ms1.90× faster
structural_objects185.8 ± 2.9 ms58.1 ± 1.6 ms30.3 ± 0.6 ms1.92× faster
interface_composition214.9 ± 41.7 ms64.7 ± 4.3 ms45.6 ± 6.0 ms1.42× faster
variadic_tuples252.9 ± 6.3 ms77.7 ± 2.5 ms42.4 ± 1.0 ms1.83× faster
checkjs_jsdoc211.5 ± 27.6 ms54.3 ± 1.2 ms36.9 ± 0.9 ms1.47× faster

Then-reported snapshot: 18/18 lower Home means; retained as history, not current admission or a restored graph speed claim. Large predicates have a 1.09× mean lead. The predeclared independent 30-round confirmation 20260827T200026Z records Home 327.1 ± 10.4 ms vs native TS 7 359.7 ± 14.2 ms (1.10× lead). The confirmation does not replace or get averaged into the full table. All 1,710 timed samples succeeded; exact versions, balanced ordering, both tables, and frozen predicate hashes were verified. All 99 program, 4,221 checker, 173 driver, and 16 harness/report tests pass, as do all shared positive projects and negative controls.

These are local synthetic results, not proof of universal, real-project, or cross-platform leadership. Full methods, retained history, and reproduction: https://github.com/home-lang/home/blob/main/docs/docs/TS_PERFORMANCE.md

Earlier 13-workload snapshot (retained)

Earlier snapshot

Apple M3 Pro, 30 interleaved runs after three warmups, mean wall time; raw result ID 20260827T085121Z. Background workstation activity increased absolute time and variance; the report retains every sample and rotates compiler order.

Workloadtsc 6.0.3native TS 7.0.2Home 0.1.0Home result
startup109.0 ms60.2 ms6.7 ms8.99× faster
many_files426.6 ms93.6 ms54.1 ms1.73× faster
deep_types260.0 ms90.3 ms28.9 ms3.12× faster
import_graph260.0 ms79.3 ms48.2 ms1.65× faster
reexport_graph195.6 ms72.6 ms57.9 ms1.25× faster
tsx_components364.1 ms90.5 ms44.5 ms2.03× faster
generic_calls360.5 ms93.8 ms51.7 ms1.81× faster
control_flow392.2 ms100.3 ms77.1 ms1.30× faster
overload_resolution491.2 ms132.3 ms88.5 ms1.49× faster
class_hierarchy406.5 ms96.0 ms83.7 ms1.15× faster
structural_objects361.8 ms102.3 ms69.4 ms1.47× faster
interface_composition532.6 ms133.1 ms122.8 ms1.08× faster
variadic_tuples634.1 ms160.8 ms153.5 ms1.05× faster

Then-accepted suite: 13/13 lower Home means.

Next coverage

  • Preserve actual global declaration types, merging and lexical isolation across files (#480)

  • Preserve generic imported/re-exported types and owner metadata, including source-order/cycle handling; readmit graph timing only after the unchanged controls pass (#487)

  • Add shared async/await and Promise type-propagation coverage with negative controls (#473; remaining #475, #476, #478)

  • Validate the legacy container path before claiming cross-platform results (#464)

  • Resolve independently tracked semantic/source-lifetime gaps (#467, #472)

  • Add a fair emit workload when output semantics are validated across all compilers

  • Add pinned real-world projects only where all compilers accept the same graph and configuration

  • Profile and resolve every regression or loss introduced by accepted workloads

  • Keep README and detailed performance documentation synchronized

Commit links and before/after measurements are maintained in the issue comments. The tracker remains open while coverage expands.

23 days ago

Follow-up probe measured and rejected: parser JSDoc-presence reuse.

  • Documentation: https://github.com/home-lang/home/commit/a6767c4e9
  • Profile: the post-gate 65,536-family sample showed 255 top-of-stack samples in the parser JSDoc diagnostic scan, so the probe reused the exact existing /** source-marker fact while preserving standalone fallback behavior.
  • Correctness: byte-identical output at 2,048, 32,768, and 65,536 families; full parser and driver suites passed.
  • Official 30-pair confirmation: candidate mean was 1.0081x lower on wall and 1.0074x lower on CPU, but both paired 95% CIs crossed zero.
  • Both 32,768-family sets were inconclusive. The first overlapped four unrelated CPU-saturating test shards; the independent set was hit by new test and service startup in rounds 7-8. Every sample, including the 4,999.780 ms and 7,011.918 ms candidate outliers, is retained and documented.
  • Decision: rejected and fully reverted because no uncontaminated retained set cleared the statistical gate. The accepted ca0bda46834cc05f9c5f57bac5dd5bdec88af388cf6b7c6fc39fcce9b60604ae binary is restored.

Raw evidence: bench/vs_tsgo/results/jsdoc-parser-presence-gate.20260901T053000Z/.

23 days ago

Measured and rejected another profile-driven probe: parser virtual-section fact reuse.

  • Documentation: https://github.com/home-lang/home/commit/f57d02a52
  • Design tested: driver-prepared sources supplied exact @filename: / @Filename: presence from the existing shared marker index; standalone parser callers retained their original discovery path.
  • Correctness: byte-identical output at 2,048, 32,768, and 65,536 predicate families; identical nonzero status and 348-byte diagnostics on a mixed virtual .d.ts/.js fixture; full parser and driver suites passed.
  • Official screen: 1.0069x lower wall and CPU means, 8/10 wins, but both paired 95% CIs crossed zero.
  • 32,768-family scale: only 1.0017x lower wall and 1.0027x lower CPU means, 4/10 wins, with both paired CIs crossing zero.
  • Decision: rejected and fully reverted. The accepted ca0bda46834cc05f9c5f57bac5dd5bdec88af388cf6b7c6fc39fcce9b60604ae binary is restored. No confirmation or competitor checkpoint was justified.

All raw samples are retained under bench/vs_tsgo/results/parser-virtual-section-fact.20260901T055116Z/.

23 days ago

Rejected and fully reverted reconciliation-sentinel root-set pruning after the fair A/B gate failed at confirmation/scale. The 2,048-family screen initially favored the candidate (wall 1.0124×, 9/10, paired CI +0.383 to +5.243 ms), but the independent 30-pair confirmation shrank to 1.0032× wall / 1.0034× CPU with both paired CIs crossing zero. At 32,768 families the candidate was slightly slower in both wall and CPU and won only 5/10 pairs. Exact output remained byte-identical at 2,048, 32,768, and 65,536 families; focused marker/gate tests passed. The accepted ca0bda… binary is restored. Documentation: https://github.com/home-lang/home/commit/2de056049 Raw evidence: bench/vs_tsgo/results/reconciliation-root-set.20260901T060254Z/

23 days ago

Rejected and fully reverted union flatten-buffer elision after its initial fair A/B gate failed. The probe avoided one temporary expansion allocation only when internUnion inputs contained no nested unions; all canonicalization semantics stayed unchanged. Exact output was byte-identical at 2,048, 32,768, and 65,536 predicate families, and focused union-interner tests passed. On the unchanged official workload, wall was 1.0056× with 6/10 candidate wins and paired CI -9.806 to +12.516 ms; CPU was 1.0035× with 6/10 and CI -8.621 to +10.312 ms. No scale/confirmation was admitted. Documentation: https://github.com/home-lang/home/commit/224d956b0 Raw evidence: bench/vs_tsgo/results/union-flatten-buffer.20260901T063000Z/

23 days ago

Rejected and fully reverted the combined prepared-parser source-fact bundle. The driver supplied exact JSDoc-comment and virtual-section presence from its existing shared marker index, while direct parser callers retained discovery fallback. Full parser/driver suites passed and exact output was byte-identical at 2,048, 32,768, and 65,536 families. The unchanged official screen remained noise: wall 1.0024×, 5/10 wins, CI -15.021 to +16.244 ms; CPU 1.0057×, 5/10, CI -8.944 to +11.802 ms. No scale or confirmation was admitted, and the accepted ca0bda… binary is restored. Documentation: https://github.com/home-lang/home/commit/8a6a285da Raw evidence: bench/vs_tsgo/results/parser-prepared-source-facts.20260901T070000Z/

23 days ago

Accepted default-export merge bucket records and pushed the source + benchmark documentation directly to main. The checker now stores one merge-entry record array per name/virtual-section bucket instead of five parallel arrays, reducing the common unique-name bucket from five allocations/growth paths to one while preserving grouping and diagnostic order. Source: https://github.com/home-lang/home/commit/1a7ac8a7a Docs: https://github.com/home-lang/home/commit/8f4fd51\n\nFair A/B results on the unchanged 2,048-family predicate workload: independent 30-pair confirmation wall 1.0285× (24/30, paired CI +3.677 to +9.625 ms) and CPU 1.0273× (25/30, CI +3.520 to +9.121 ms). At 32,768 families CPU is 1.0238× (9/10, CI +10.202 to +173.825 ms); wall is lower but explicitly inconclusive. Both 65,536 means are lower but their intervals cross zero and remain labeled inconclusive. Exact output is byte-identical at all three scales; full ReleaseFast checker and 95/95 harness tests pass. Raw A/B: bench/vs_tsgo/results/default-export-bucket-records.20260901T074000Z/.\n\nPinned schema-3 checkpoints compare JavaScript TypeScript 6.0.3 with native TypeScript/tsgo 7.0.2. Focused 20260901T065015Z: tsc 1004.8 ± 26.3 ms, tsgo 347.4 ± 9.4 ms, Home 224.5 ± 10.1 ms (1.55× vs TS7; 30/30 wins; CI +118.439 to +127.422 ms). Complete 20260901T065201Z: Home has the lowest mean on 20/20 workloads, wins 600/600 TS7 pairs, and every paired CI is positive; narrowest lead 1.53×.

23 days ago

Documented and pushed the rejected type-alias merge inline-first target probe in c0a5079: https://github.com/home-lang/home/commit/c0a507942

The exact representation probe kept the first class/interface target inline and allocated overflow storage only for true multi-target groups. Correctness and exact-output checks passed at 2,048, 32,768, and 65,536 predicate families, but the official 10-pair screen was inconclusive: wall 1.0089x with 6/10 wins and CI [-5.243, +9.348] ms; CPU 1.0049x with 5/10 wins and CI [-4.146, +6.371] ms. Two unrelated Zig test shards imposed a stable two-core external load throughout; all samples were retained. The initial gate failed, so no scale or competitor checkpoint was admitted.

The source probe was fully reverted and the accepted binary (SHA-256 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d) restored. Raw evidence: bench/vs_tsgo/results/type-alias-merge-inline-first.20260901T083000Z/.

23 days ago

Documented and pushed the rejected duplicate-class threshold flag probe in 483baca: https://github.com/home-lang/home/commit/483bacad5

The exact representation probe replaced an unnecessary ClassGroup usize count with a boolean threshold while preserving the first declaration anchor and diagnostics. Focused three-class/two-interface coverage, the complete ReleaseFast ts_checker suite, and byte-exact empty output at 2,048, 32,768, and 65,536 predicate families all passed.

The first 10-pair set was retained but excluded because its load snapshots showed an unrelated Bun job at the start that had exited by the end. A predeclared rerun with the same frozen binaries was directionally slower: wall 226.580 +/- 2.854 ms baseline vs 228.710 +/- 4.979 ms candidate (1.0094x slower, 4/10 wins, CI [-4.636, +0.375] ms); CPU 225.118 +/- 2.760 vs 227.103 +/- 4.789 ms (1.0088x slower, 4/10 wins, CI [-4.247, +0.278] ms). A separate test262 shard and Spotlight each held approximately one core through the rerun; every sample was retained. The gate failed, so no scale timing or TS 6.0.3 vs native TS 7.0.2 checkpoint was admitted.

The probe and regression are fully reverted and the accepted binary SHA-256 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d is restored. Raw evidence: bench/vs_tsgo/results/duplicate-class-flag.20260901T091500Z/.

23 days ago

Documented and pushed the rejected duplicate-class second-pass gate in 714ad6c: https://github.com/home-lang/home/commit/714ad6cf8

The exact necessary-condition probe skipped the interface rescan only when the complete class scan found no duplicate class group. Focused three-class/two-interface coverage, the complete ReleaseFast ts_checker suite, and byte-exact empty output at 2,048, 32,768, and 65,536 predicate families all passed.

The official 10-pair screen did not clear admission: wall 225.297 +/- 3.702 ms baseline vs 223.708 +/- 2.394 ms candidate (1.0071x, 5/10 wins, CI [-0.394, +3.571] ms); CPU 223.789 +/- 3.469 vs 222.307 +/- 2.316 ms (1.0067x, 7/10 wins, CI [-0.355, +3.319] ms). A separate one-core zig-js benchmark was present in both load snapshots; its worker changed while the load class remained present, and every sample was retained. Both intervals cross zero, so no confirmation, scale timing, or TS 6.0.3 vs native TS 7.0.2 checkpoint was admitted.

The probe and regression are fully reverted and the accepted binary SHA-256 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d is restored. Raw evidence: bench/vs_tsgo/results/duplicate-class-second-pass-gate.20260901T094500Z/.

23 days ago

Documented and pushed the rejected local-boundary NodeId key probe in 0bc051c: https://github.com/home-lang/home/commit/0bc051c5d

The exact representation probe removed redundant virtual-section data from statement-boundary map keys while keeping declaration keys section-partitioned. The exhaustive indexed-vs-slow lookup suite (multiple virtual files, export wrappers, nested scopes, allocation failure), complete ReleaseFast ts_checker suite, and byte-exact empty output at 2,048, 32,768, and 65,536 predicate families all passed.

The official 10-pair screen was a near tie and failed admission: wall 228.922 +/- 3.841 ms baseline vs 228.610 +/- 2.939 ms candidate (1.0014x, 6/10 wins, CI [-1.605, +2.229] ms); CPU 227.242 +/- 3.518 vs 227.108 +/- 2.872 ms (1.0006x, 6/10 wins, CI [-1.398, +1.665] ms). A separate one-core zig-js parent/candidate benchmark remained present in both load snapshots; every sample was retained. No confirmation, scale timing, or TS 6.0.3 vs native TS 7.0.2 checkpoint was admitted.

The probe is fully reverted and the accepted binary SHA-256 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d restored. Raw evidence: bench/vs_tsgo/results/local-boundary-node-keys.20260901T102500Z/.

23 days ago

Documented and pushed the rejected namespace declaration-pass gate in 126659a: https://github.com/home-lang/home/commit/126659a8c

The exact necessary-condition probe used the existing namespace/module/global source marker to skip four namespace-only declaration scans, while false positives and HIR-only/source-less integrations retained the original path. Focused namespace/ambient-module coverage, an explicit HIR-only fallback regression, the complete ReleaseFast ts_checker suite, and byte-exact empty output at 2,048, 32,768, and 65,536 predicate families all passed.

The first screen was retained but excluded because a 116% Bun job disappeared while Time Machine began across its load snapshots. The predeclared clean rerun still failed admission: wall 218.447 +/- 2.484 ms baseline vs 218.034 +/- 2.028 ms candidate (1.0019x, 6/10 wins, CI [-1.380, +2.207] ms); CPU 217.430 +/- 2.401 vs 216.850 +/- 1.979 ms (1.0027x, 7/10 wins, CI [-1.109, +2.269] ms). No samples were filtered, and no confirmation, scale timing, or TS 6.0.3 vs native TS 7.0.2 checkpoint was admitted.

The probe and regression are fully reverted and the accepted binary SHA-256 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d restored. Raw evidence: bench/vs_tsgo/results/namespace-declaration-pass-gate.20260901T111500Z/.

23 days ago

Rejected and documented the recovered-parameter identity-reuse probe in a62a4ff.

  • Baseline: 9e86ee995ca30484ae41d0f0adafea23a386d1ad57fd932b4941909c20d4711d
  • Candidate: b7b91c25a92f5835b224425e4f16d7370bc532cc34e87c2a7e5fdcdbb8f9a10d
  • Exact empty output at 2,048 / 32,768 / 65,536 predicate families; focused + full checker and 95/95 harness tests passed.
  • 10-pair official screen favored the candidate (~0.56% wall, ~0.59% CPU), but the independent 30-pair confirmation crossed zero for both metrics.
  • The 65,536-family scaling run also crossed zero for wall and CPU despite 8/10 paired wins.

No inconclusive micro-optimization was admitted: source reverted, accepted binary restored. Raw evidence remains under bench/vs_tsgo/results/recovered-parameter-identity-reuse.20260901T123000Z/. The benchmark manifest and installed native compiler remain pinned and verified at TypeScript/tsgo 7.0.2; TS 7 and tsgo are one competitor, compared against JavaScript TypeScript 6.0.3.

23 days ago

Accepted and published the token-count HIR reservation optimization.

Commits:

  • Source: ed8bf949b — perf(ts): reserve HIR storage from token count (#416)
  • Benchmarks/docs: f58aa516c — docs(ts): publish HIR reservation benchmarks (#416)

What changed: after lexing determines the exact token count, the parser reserves proportional capacity for the five parallel HIR node columns and the child pool. This changes capacity only; recovery can still grow normally.

Correctness gates:

  • complete HIR, TS driver, TS parser, and ReleaseFast ts_checker suites passed
  • focused HIR invariant test passed
  • zig fmt --check, git diff --check, and all 95 benchmark-harness tests passed
  • baseline/candidate stdout and stderr were byte-identical at 2,048, 32,768, and 65,536 predicate families

Fair alternating-order A/B:

  • official 2,048-family, independent 30-pair confirmation: wall 1.0189× faster (26/30 wins, paired 95% CI +1.456 to +7.648 ms); CPU 1.0163× faster (27/30, +1.420 to +6.333 ms)
  • diagnostic 32,768-family, 10 pairs: wall 1.0067× faster (8/10, +1.671 to +50.651 ms); CPU 1.0057× faster (8/10, +0.138 to +44.724 ms)
  • official peak RSS increased by 0.509 MiB; at 32,768 families mean RSS decreased by 3.147 MiB. All samples were retained.

Fresh schema-3 competitor checkpoints use exactly tsc 6.0.3 versus native TypeScript/tsgo 7.0.2 (TS 7 and tsgo are the same competitor):

  • focused type_predicates_large: Home 221.7 ± 6.0 ms versus native TS 7.0.2 at 350.3 ± 11.1 ms (1.58× faster), 30/30 paired wins
  • full 20-workload checkpoint: Home has the lowest mean on 20/20 and wins 600/600 paired wall rounds against native TS 7.0.2; every workload's paired interval is positive. The narrowest mean lead is 1.59×.

Detailed evidence and raw-result identifiers are in TS_PERFORMANCE.md; the README contains the brief linked table.

23 days ago

Profiled and rejected the checker-directive marker indexing probe; no source change landed.

Documentation commit:

  • 2884a46cd — docs(ts): record directive marker probe rejection (#416)

Probe: register the exact allowUnreachableCode and noImplicitThis spellings in the existing one-pass source-marker index, replacing three independent post-check whole-source searches. The post-change profile confirmed those three scans disappeared.

Correctness:

  • complete ReleaseFast checker and TS driver suites passed
  • focused noImplicitThis: false behavior passed
  • all 95 benchmark-harness tests passed
  • zig fmt --check and git diff --check passed
  • baseline/candidate output was byte-identical at 2,048 and 32,768 predicate families

Exact final-source official evidence, all samples retained:

  • first 30 pairs: wall 1.0052× (20/30), CI -0.348 to +2.503 ms; CPU 1.0051× (20/30), CI -0.327 to +2.446 ms
  • second independent 30 pairs: wall 1.0198× (24/30), CI +1.415 to +8.487 ms; CPU 1.0179× (25/30), CI +1.520 to +7.024 ms
  • precommitted aggregate of all 60 final-binary pairs: wall 1.0125× (44/60), CI +1.070 to +4.996 ms; CPU 1.0115× (45/60), CI +1.109 to +4.274 ms; RSS statistically flat

Required final-source 32,768-family scale gate:

  • wall 1.0203× lower mean (7/10), but paired CI -416.122 to +760.379 ms crosses zero
  • CPU 1.0418× lower mean (7/10), CI +0.428 to +550.640 ms
  • the retained host snapshots show a one-core Zig build at the start and ten unrelated zig-js tests consuming roughly 80–91% of a core each at the end; paired wall differences ranged from -1473.438 to +2517.279 ms

No sample was filtered, replaced, or selectively rerun. Because the precommitted scale wall gate remained inconclusive, I rejected and fully reverted the probe despite the positive 60-pair official aggregate. Raw evidence is retained locally under bench/vs_tsgo/results/directive-marker-index.20260901T090828Z/; the detailed public record is in TS_PERFORMANCE.md, with a brief README table row.

Compiler comparison pins remain exactly tsc 6.0.3 and native TypeScript/tsgo 7.0.2; this rejected internal A/B did not replace the current admitted competitor snapshot.

23 days ago

Rejected and documented the follow-up broad checker-directive marker gate; no source change landed.

Documentation commit:

  • 3751fe1cf — docs(ts): record broad directive gate rejection (#416)

This smaller variant did not enlarge the source-marker automaton. It used the already-indexed broad @allow marker to skip the two exact allowUnreachableCode searches when absent, and the already-indexed @ marker to skip the exact noImplicitThis search when absent. Files containing those broad markers retained the original exact searches and semantics.

Correctness:

  • complete ReleaseFast checker suite passed
  • focused noImplicitThis: false behavior passed
  • zig fmt --check and git diff --check passed
  • frozen baseline/candidate output was byte-identical and empty on the unchanged official 2,048-family project

Primary fair A/B after three alternating warmup pairs, with reversed order and all 10 pairs retained:

  • wall: baseline 214.961 ± 1.612 ms; candidate 215.475 ± 3.620 ms; 0.9976× baseline/candidate, 4/10 candidate wins; paired CI -2.748 to +1.359 ms
  • CPU: baseline 213.774 ± 1.480 ms; candidate 214.255 ± 3.497 ms; 0.9978×, 6/10 candidate wins; paired CI -2.542 to +1.271 ms
  • RSS differed by about 0.003 MiB in the candidate's direction

The candidate has slightly worse wall and CPU means, and both timing intervals cross zero. It therefore failed the primary admission gate and was fully reverted without a confirmation, scale run, source commit, or competitor checkpoint. One unrelated Typesense process and Spotlight were active in the retained load snapshots; no sample was filtered or rerun.

Raw evidence is retained locally under bench/vs_tsgo/results/directive-broad-marker-gate.20260901T094320Z/. Detailed results are in TS_PERFORMANCE.md, with a brief README table row.

23 days ago

Rejected and fully reverted the stable pool-index inference probe after its independent confirmation failed the precommitted timing gate.

  • Logical change tested: replace recursive-inference member-slice copies with one-value-at-a-time reads through stable pool offsets.
  • Correctness: focused recursive-inference regression and complete ReleaseFast ts_checker suite passed; official baseline/candidate outputs were byte-identical.
  • 10-pair official screen: 1.0104x wall and 1.0100x CPU; 10/10 wins; both paired 95% CIs positive.
  • Independent 30-pair confirmation: 1.0150x wall mean (20/30 wins, CI -1.575 to +8.838 ms) and 1.0101x CPU mean (20/30 wins, CI -1.203 to +5.471 ms).
  • Every sample was retained. Because both confirmation timing intervals cross zero, no scale run, source commit, or TS 6.0.3 vs native TS 7.0.2 competitor checkpoint was admitted.

Documentation: cb63644 Raw evidence: bench/vs_tsgo/results/inference-stable-pool-index.20260901T095901Z/

23 days ago

Rejected and fully reverted the primitive type-name dispatch probe after its required scale gate remained inconclusive.

  • Logical change tested: route unresolved-signature primitive checks through the existing classifier and dispatch the exact same 12 lowercase spellings by length/first byte.
  • Correctness: focused regression, complete ReleaseFast ts_checker and TS driver suites, 95/95 benchmark-harness tests, formatting, and exact-output gates all passed.
  • Official 10-pair screen: 1.0046x wall and 1.0042x CPU; both paired 95% CIs positive.
  • Independent 30-pair confirmation: 1.0218x wall (20/30 wins, CI +1.983 to +8.664 ms) and 1.0200x CPU (22/30 wins, CI +2.150 to +7.167 ms).
  • Required 32,768-family set: 1.0027x wall and 1.0057x CPU, but both paired CIs crossed zero (-74.299 to +74.560 ms wall; -42.135 to +74.755 ms CPU).
  • Every sample was retained. Load snapshots record the unrelated queued threadfuzz and later Typesense/macOS activity. No source commit or TS 6.0.3 vs native TS 7.0.2 checkpoint was admitted.

Documentation: 1541b1b Raw evidence: bench/vs_tsgo/results/primitive-type-name-dispatch.20260901T102051Z/

23 days ago

Published the free-type-parameter root-memo rejection in e6947c3.

Official 2,048-family screen (10 alternating A/B pairs):

  • wall: baseline 216.023 ± 2.201 ms, candidate 217.643 ± 4.010 ms (0.9926×; 4/10 wins; paired 95% CI -3.836 to +0.372 ms)
  • CPU: baseline 214.927 ± 2.140 ms, candidate 216.599 ± 4.005 ms (0.9923×; 4/10 wins; paired 95% CI -3.795 to +0.246 ms)
  • peak RSS: baseline 75.139 ± 0.060 MiB, candidate 75.197 ± 0.015 MiB (0.9992×; 0/10 wins; paired 95% CI -0.094 to -0.022 MiB)

The candidate was slower and used more memory in every pair, so it was rejected before confirmation/scale. Source and focused regression are fully reverted. The detailed methodology, load disclosure, hashes, and raw-result path are documented in docs/docs/TS_PERFORMANCE.md; the README summary table is updated too.

No source commit or competitor checkpoint was admitted.

23 days ago

Published the free-type traversal generation-mark rejection in 0b20a4c.

The u32 form failed its 10-pair screen: timing CIs crossed zero and RSS was conclusively higher. A byte-epoch refinement then passed its screen (1.0227× wall and 1.0222× CPU, 10/10 wins, positive paired intervals) without an RSS penalty.

The required independent 30-pair confirmation did not establish the timing gain:

  • wall: 219.443 ± 3.498 ms baseline vs 218.508 ± 6.756 ms candidate (1.0043×; 23/30 wins; paired 95% CI -1.382 to +2.543 ms)
  • CPU: 218.265 ± 3.408 ms vs 217.191 ± 6.019 ms (1.0049×; 23/30 wins; CI -0.868 to +2.469 ms)
  • peak RSS: 75.080 ± 0.011 MiB vs 74.975 ± 0.111 MiB (CI +0.069 to +0.142 MiB)

Because both timing intervals cross zero, no scale run was admitted. Both source forms and the rollover regression are fully reverted. The detailed report and raw-result paths are in docs/docs/TS_PERFORMANCE.md, with a brief README table row. No source commit or competitor checkpoint was admitted.

23 days ago

Accepted and published the contextual cache descendant index.

Source: 21e6155 (perf(ts): index contextual cache descendants) Docs/README: d205881

The retained 4,096-family interface-composition profile showed clearCachedTypesWithin as the largest exclusive leaf (1,018 samples). The accepted implementation lazily indexes exact HIR parent→children edges and clears only the contextual function subtree; allocation failure falls back to the unchanged whole-HIR scan.

Official 128-family independent confirmation (30 reversed-order pairs):

  • wall: 39.353 ± 0.657 ms → 37.441 ± 0.684 ms (1.0511×; 29/30 wins; paired 95% CI +1.547 to +2.268 ms)
  • CPU: 38.498 ± 0.590 ms → 36.582 ± 0.584 ms (1.0524×; 30/30 wins; CI +1.618 to +2.203 ms)
  • peak RSS: 16.422 ± 0.008 MiB → 16.387 ± 0.036 MiB (CI +0.022 to +0.047 MiB)

Required 16× scale, 2,048 families (10 pairs):

  • wall: 3913.685 ± 89.745 ms → 3397.317 ± 35.948 ms (1.1520×; 10/10 wins; CI +459.034 to +564.619 ms)
  • CPU: 3903.645 ± 85.059 ms → 3389.624 ± 35.453 ms (1.1516×; 10/10 wins; CI +459.204 to +559.232 ms)
  • peak RSS rises by 1.5 MiB at this 92,160-line scale, the disclosed cost of two u32 index entries per HIR node.

Fresh targeted 30-run compiler checkpoint (TS 6.0.3 vs native TS 7.0.2 vs Home): 202.1 ± 6.3 ms, 66.1 ± 9.9 ms, and 40.8 ± 18.7 ms respectively. Home is 1.62× faster than TS7 with 29/30 wins and a +21.383 to +27.862 ms paired interval. The 139.8 ms first Home outlier remains included.

Correctness: focused descendant/sibling/index-reuse regression; complete ReleaseFast checker, driver, and CLI suites; all 20 version-pinned workload admissions and existing negative controls. Exact baseline/candidate output matches on official, 2,048-family, and 4,096-family projects.

All raw rounds, binaries, hashes, load snapshots, provenance, and corpora are linked from docs/docs/TS_PERFORMANCE.md. The README summary table is updated.

23 days ago

Published the next accepted TypeScript frontend optimization in 8d1d198b2, with docs/README in c4e5775b0.

The checker now uses complementary exact declaration-name indexes to prove when resolveValueDeclInStmt cannot succeed, skipping the full statement scan only for ordinary blocks with at least 16 statements. Virtual sections, recovery HIR, incomplete indexes, and allocation failures retain the original scan.

Final same-correctness A/B against accepted 21e6155d4:

  • official 128 families, 30 pairs: wall 36.772 -> 34.771 ms (1.0575x, 30/30 wins, paired CI +1.644..+2.348 ms); CPU 35.991 -> 33.947 ms (1.0602x, 30/30, CI +1.738..+2.343 ms); RSS +0.030 MiB
  • independent 2,048-family / 92,160-line scale, 10 pairs: wall 3285.301 -> 2621.967 ms (1.2530x, 10/10, CI +624.290..+703.416 ms); CPU 1.2525x; RSS +0.036 MiB
  • every sample retained; screen/confirmation/scale use reversed order after three alternating warmup pairs

All 20 positive and negative admissions pass against JavaScript TS 6.0.3 and the single native TS/tsgo 7.0.2 competitor. Final 30-round checkpoint: TS6 219.7±13.9 ms, native TS7 76.0±28.0 ms, Home 47.1±38.7 ms; Home 1.61x faster, 29/30 paired wins, TS7-minus-Home CI +16.520..+38.149 ms. Two early Home outliers remain included.

Verification: checker 4,325/4,325, driver 188/188, CLI 69/69, benchmark harness 95/95, Pickier clean. Raw final evidence: bounded-complementary-value-index.20260901T121547Z/, admission 20260901T122448Z/, checkpoint 20260901T122547Z/. The two memory-heavier refinements were rejected and documented separately.

23 days ago

Published another accepted profile-backed frontend optimization in b3dcb63c5, with detailed docs and the linked README row in 02b131431.

Fresh accepted-binary profiling showed virtualExportAssignmentTargetDeclForLocal taking 1,015 exclusive samples while scanning import-free roots. The helper now returns from the existing exact negative import-declaration fact; import-capable sources and HIR-only callers keep the original lookup.

Same-correctness A/B versus accepted 8d1d198b2:

  • official confirmation, 30 pairs: wall 36.161 -> 34.845 ms (1.0378x, paired CI +0.833..+1.802 ms); CPU 1.0377x, CI +0.898..+1.663 ms; RSS CI crosses zero
  • 2,048-family / 92,160-line scale, 10 pairs: wall 2728.437 -> 2351.781 ms (1.1602x, 10/10 wins, CI +349.128..+404.735 ms); CPU 1.1597x; RSS CI crosses zero
  • every sample retained with alternating order and before/after load snapshots

All 20 positive and negative admissions pass against JavaScript TS 6.0.3 and the single native TS/tsgo 7.0.2 competitor. Final 30-round checkpoint: TS6 210.3±8.6 ms, native TS7 67.5±11.8 ms, Home 35.7±2.0 ms; Home 1.89x faster, 30/30 paired wins, TS7-minus-Home CI +29.141..+36.480 ms.

Verification: checker 4,325/4,325, driver 188/188, CLI 69/69, benchmark harness 95/95, Pickier clean. Evidence: interface-composition-profile-final.20260901T122859Z/, import-free-export-assignment-gate.20260901T123335Z/, admission 20260901T123630Z/, checkpoint 20260901T124307Z/.

23 days ago

Recorded and reverted the next profile experiment in 384fe05c1.

The candidate reused the exact HIR namespace index to skip globalAugmentedValueType root scans when no decoded global namespace existed. It preserved virtual sections, escaped identifiers, multiple matches, and allocation-failure fallback, but failed the fair official screen:

  • wall 35.809 -> 35.962 ms (0.9957x, 4/10 wins, paired CI -1.491..+1.072 ms)
  • CPU 34.662 -> 34.560 ms (1.0030x, 6/10 wins, CI -1.017..+1.036 ms)
  • RSS was slightly lower, but timing was inconclusive and wall mean regressed

Every sample is retained under global-namespace-negative-index.20260901T124843Z/. The source is fully reverted; no confirmation, scale timing, source commit, or competitor checkpoint was admitted. README and detailed docs both link the rejection so it cannot be mistaken for an accepted win.

23 days ago

Rejected and fully reverted the qualified-annotation root namespace index probe.

  • Profile-backed target: typeAnnotationRootIsUnresolved repeatedly rescanned the source root for one-segment namespaces.
  • Exact candidate: reused the completed namespace index for ordinary sources, preserved scans for virtual sections/incomplete indexes, and skipped the import loop only when the existing conservative source marker proved no import token.
  • Baseline SHA-256: f451419a6bbd8e897e9f0b048f992a8034866d95d458591c916f090be0a95e23
  • Candidate SHA-256: 6b3e6d85483a1124ebdc6a2b4547c8d68752c40ef3e7241e6dcbafda628b2548
  • Official and 16x outputs were byte-identical and empty.
  • 10-pair official screen retained every sample: wall 52.427 -> 48.469 ms (1.0817x, 6/10, CI -0.642..+9.406 ms); CPU 50.360 -> 48.613 ms (1.0359x, 7/10, CI -0.635..+4.213 ms); RSS essentially flat.
  • Ten unrelated Zig test shards occupied most cores in both snapshots (load 26.33 -> 29.23). No rounds were filtered or rerun. Both timing intervals crossed zero, so no confirmation, scale timing, correctness suite, source commit, or TS 6.0.3 vs native TS/tsgo 7.0.2 checkpoint was admitted.

Documentation: https://github.com/home-lang/home/commit/588ee98cb Raw local evidence: bench/vs_tsgo/results/qualified-type-root-namespace-index.20260901T125755Z/

23 days ago

Accepted and pushed the next profile-backed optimization: the namespace-value miss index.

Source: https://github.com/home-lang/home/commit/4ba46f4d6 Documentation and README table: https://github.com/home-lang/home/commit/dee0937e9

What changed:

  • mergedNamespaceValueMemberType now returns an exact miss after its visible-namespace lookup has already completed the root index and found no namespace.
  • Virtual filename sections, visible/reopened namespaces, dotted namespace roots, and incomplete indexes retain the original fallback scan.
  • Dotted-marker allocation failure now leaves the namespace index incomplete, preserving the scan.

Frozen provenance:

  • Baseline SHA-256: f451419a6bbd8e897e9f0b048f992a8034866d95d458591c916f090be0a95e23
  • Candidate SHA-256: 4c0f758c9c8a8bf313ced6fd4d8b872e6f6b96166503b7c4035e8141a2fea7a9
  • Official and 16x outputs: byte-identical, empty, exit zero.

Fair A/B results, every pair retained, alternating order after three warmup pairs:

  • Official screen (10): wall 34.655 -> 33.321 ms, 1.0400x, CI +0.486..+2.185; CPU 33.709 -> 32.306 ms, 1.0434x, CI +0.720..+2.092; RSS flat.
  • Official confirmation (30): wall 36.506 -> 35.183 ms, 1.0376x, CI +0.700..+1.937; CPU 35.480 -> 34.027 ms, 1.0427x, CI +0.977..+1.930; RSS flat.
  • 2,048-family scale (10): wall 2506.140 -> 2077.577 ms, 1.2063x, 10/10, CI +387.978..+479.160; CPU 2486.218 -> 2057.291 ms, 1.2085x, 10/10, CI +389.549..+479.714; RSS flat.

Correctness:

  • Focused dotted- and merged-namespace tests pass.
  • ReleaseFast checker 4,325/4,325, driver 188/188, CLI 69/69.
  • Benchmark harness 95/95.
  • All 20 workload admissions and rejection controls pass with verified TS 6.0.3 and native TS/tsgo 7.0.2 provenance (20260901T132853Z).

Final 30-round checkpoint (20260901T132947Z): TS 6.0.3 210.2 +/- 5.2 ms; native TS 7.0.2 70.8 +/- 19.9 ms; Home 34.7 +/- 5.1 ms. Home is 2.04x faster than native TS7 with 30/30 paired wins and TS7-minus-Home CI +30.368..+44.490 ms. The first 61.2 ms Home outlier remains included.

Raw local A/B evidence: bench/vs_tsgo/results/namespace-value-miss-index.20260901T130826Z/.

23 days ago

Accepted the enum-free visible-namespace index in two small direct-to-main commits:

The fresh 4,096-family profile selected visibleTypeDeclarationExistsAt (776 exclusive samples). The implementation reuses the exact per-container namespace index only for attached, enum-free, nonvirtual sources; incomplete indexes, enum-bearing sources, virtual sections, allocation failure, and source-less HIR keep the original scan.

Fair retained A/B results:

  • Official 128-family screen, 10 pairs: 1.0265x wall and 1.0296x CPU; paired 95% CIs +0.312..+1.423 ms and +0.520..+1.352 ms.
  • Independent official confirmation, 30 pairs: 1.0318x wall and 1.0323x CPU; 25/30 wins for both; paired CIs +0.505..+1.627 ms and +0.596..+1.491 ms.
  • Independent 2,048-family scale, 10 pairs: 1.1308x wall and 1.1315x CPU; 10/10 wins; paired CIs +215.583..+271.102 ms and +218.572..+266.353 ms.
  • Scale RSS increased from 114.320 to 114.755 MiB (+0.435 MiB mean; baseline-minus-candidate CI -0.697..-0.173 MiB). This tradeoff is explicitly documented.

A 16-statement bounded refinement was rejected: both screen timing CIs crossed zero, so it received no confirmation or scale run. Its samples remain separate.

Correctness: ReleaseFast checker 4,325/4,325, driver 188/188, CLI 69/69, harness 95/95, plus all 20 version-pinned admissions and controls.

Final unfiltered 30-round competitor checkpoint:

  • TypeScript 6.0.3: 220.1 +/- 46.4 ms
  • Native TypeScript/tsgo 7.0.2: 69.9 +/- 21.6 ms
  • Home: 35.5 +/- 9.2 ms, 1.97x faster, 29/30 wins, paired TS7-minus-Home CI +29.691..+40.983 ms

All samples, including the shared outlier round, remain retained. TS 7 and tsgo are treated as the same single native competitor throughout.

23 days ago

Rejected and fully reverted the next profile-guided probe: unique interface-heritage index.

Documentation: https://github.com/home-lang/home/commit/86578ff2f

A fresh independent 8,192-family profile confirmed visibleTypeDeclarationExistsAt had fallen to one residual sample after the accepted optimization. The next untested leaf was interfaceHeritageReachesName at 1,029 exclusive samples; globalAugmentedValueType was skipped because its exact negative-index form had already failed admission.

The candidate used a pass-local exact map for unique interface names. Duplicate names, small statement lists, and incomplete allocation retained the original scan. Both frozen binaries produced identical empty output on the official and 2,048-family scale projects, and the ReleaseFast checker suite passed.

Official 128-family screen, all 10 reversed-order pairs retained:

  • Wall: 35.993 +/- 1.751 ms baseline vs 36.162 +/- 1.586 ms candidate; 0.9953x, 4/10 wins, paired CI -1.142..+0.824 ms.
  • CPU: 34.711 +/- 1.509 ms baseline vs 34.792 +/- 1.436 ms candidate; 0.9977x, 4/10 wins, paired CI -0.933..+0.859 ms.
  • RSS: 16.608 +/- 0.009 MiB baseline vs 16.591 +/- 0.019 MiB candidate; paired CI +0.005..+0.031 MiB.

The slight RSS reduction does not override the timing rule: both timing means regressed and both CIs crossed zero. No confirmation, scale timing, source commit, or TS 6.0.3 vs native TS/tsgo 7.0.2 checkpoint was admitted. Raw evidence remains separate under interface-heritage-index.20260901T140413Z/.

23 days ago

Published 2b4de1114 documenting two exact merged-interface scan-index designs, both rejected and fully reverted. The broad clean screen was 1.0146× wall / 1.0155× CPU, but paired CIs crossed zero (-0.770..+1.695 ms wall; -0.640..+1.612 ms CPU). The refined duplicate-only chain was 0.9990× wall / 1.0024× CPU, again with crossing CIs (-0.901..+0.871; -0.518..+0.754). Full checker correctness and byte-identical official/scale outputs passed for both. README summary and detailed evidence/hashes are linked from the performance docs.

23 days ago

Published 5615aab14 documenting two namespace merge-order probes, both rejected and fully reverted. The exact last-ordinal index screened at 1.0134× wall / 1.0110× CPU, but both paired CIs crossed zero (-0.319..+1.288 ms; -0.261..+1.041 ms). A simpler reuse of the existing function index regressed to 0.9889× wall / 0.9954× CPU with crossing CIs. Both passed the full ReleaseFast checker suite and byte-identical official/scale outputs. No confirmation, scale timing, source commit, or checkpoint was admitted.

23 days ago

Published ed4696540 documenting the rejected declarationName forced-inline probe. The clean screen was 1.0104× wall / 1.0060× CPU in mean, but paired CIs crossed zero (-0.138..+0.931 ms; -0.128..+0.574 ms). The candidate also increased the binary by 16,544 bytes (0.135%) and measured RSS. Full checker correctness and byte-identical official/scale outputs passed; the one-token source change was fully reverted. A visibly loaded preliminary set is retained separately and excluded.

23 days ago

Accepted and published the import-free virtual-import gate.

  • Source: 83c44368c
  • Benchmark docs + README summary: c41140b11
  • TypeScript comparison is pinned to JS TypeScript 6.0.3 versus the single native TypeScript/tsgo 7.0.2 competitor.
  • Official 128-family confirmation: 1.0327x wall, 18/30 wins, paired CI +0.138 to +2.649 ms.
  • 2,048-family scale: 1.0576x wall, 10/10 wins, paired CI +70.251 to +129.207 ms.
  • Final competitor checkpoint: Home 33.2 ms versus native TS 7.0.2 at 69.5 ms; 2.09x faster, 30/30 paired wins.

All 20 admission controls and the checker, driver, CLI, and 95-test benchmark harness passed. Full methodology, retained outliers, hashes, and artifacts are documented.

23 days ago

Starting the next profile-backed TypeScript frontend probe from clean origin/main (37215a5f7). A fresh 10-second sample of the independently scaled 8,192-family interface_composition project (368,640 generated lines; unchanged generator and options) records 271 exclusive samples in scriptObjectExpandoHasPrototypeAssignment. That helper scans the whole source root for NS.Member.prototype assignments even when the checker’s existing exact conservative .prototype marker proves the match impossible. I am evaluating a one-line attached-source negative gate; source-less HIR and every source containing .prototype retain the existing scan. Admission requires byte-identical outputs, a positive paired official screen and independent confirmation, the 16x scale check, correctness suites, and the version-pinned TS 6.0.3/native TS 7.0.2 controls. Any inconclusive or regressing candidate will be fully reverted and documented.

23 days ago

Rejected and fully reverted the profile-backed prototype-assignment marker gate; documentation and the linked README row are now on main in abd00d1d2. The initial loaded set is retained (32.053 -> 35.239 ms wall; paired CI -8.687..+2.315 ms). After the identified transient processes exited, the separately labeled 10-pair replacement measured 31.648 -> 31.780 ms wall (0.9958x, 5/10 wins, CI -0.813..+0.549) and 30.695 -> 30.781 ms CPU (0.9972x, 5/10, CI -0.792..+0.621). Both binaries were the same size and produced byte-identical empty official output. Because both timing intervals crossed zero and the means regressed, there was no confirmation, scale run, correctness suite, source commit, or competitor checkpoint. Every sample, binary, profile, hash, exact output, and load snapshot remains under script-expando-prototype-gate.20260902T211936Z/.

23 days ago

Starting the next independent profile-backed probe after the fully reverted prototype gate. The same accepted-code 8,192-family trace records 236 exclusive samples in moduleNamespaceTypeForLocalImport, the next untested whole-root miss scan. That helper can only match an import declaration, but unlike the already accepted virtualImportTypeForLocal gate it does not consult the existing conservative source_may_have_import_declaration fact. I am evaluating the corresponding attached-source negative gate; import-bearing source and source-less HIR retain the complete existing module/CommonJS/interop resolution path. The frozen accepted binary remains the baseline, and the same screen/confirmation/scale/correctness/admission rules apply.

23 days ago

Rejected and fully reverted the import-free module-namespace gate; documentation and the README link are pushed to main in 3cfaa51e6. The unchanged official 128-family screen retained every pair: wall 33.200 -> 33.275 ms (0.9978x, 5/10 wins, paired CI -0.816..+0.667) and CPU 31.856 -> 31.931 ms (0.9977x, 6/10, CI -0.575..+0.425). Both binaries were 12,026,760 bytes and produced byte-identical empty output. Two unrelated Zig builds occupied two full cores in both snapshots; since the candidate still regressed, there was no replacement, confirmation, scale timing, correctness suite, source commit, or competitor checkpoint. All evidence is retained under import-free-module-namespace-gate.20260902T212617Z/.

23 days ago

Starting the next independent profile target: the accepted 8,192-family trace records 211 exclusive samples in findTypeAliasDeclForHeritageCycle. Its first findTypeAliasDeclInScope lookup already consults the existing conservative type-alias source fact, but after an exact syntax miss the heritage fallback still rescans every root statement for a type alias. I am evaluating reuse of that same negative fact before only the fallback scan. Type-alias-bearing sources and source-less HIR retain the complete scoped and root lookup. The unchanged official corpus and all established admission rules remain the gate.

23 days ago

Rejected and fully reverted the type-alias heritage fallback gate; detailed evidence and the README row are pushed to main in 5d03cf219. Loaded screen: wall 34.963 -> 34.010 ms (1.0280x, 7/10 wins, paired CI -0.406..+2.311) and CPU 1.0278x with a crossing interval. After the two identified links exited, the one predeclared replacement encountered ten unrelated test shards and a new link; it retained every pair and measured wall 39.621 -> 38.878 ms (1.0191x, 6/10, CI -1.744..+3.229) and CPU 1.0112x with a crossing interval. I did not repeat measurements until noise favored the candidate. Byte-identical official output and equal 12,026,760-byte binaries passed; no confirmation, scale timing, correctness suite, source commit, or competitor checkpoint was admitted. Evidence remains under type-alias-heritage-fallback-gate.20260902T213124Z/.

23 days ago

Progress via #548: pushed https://github.com/home-lang/home/commit/3852cd38fb88d9e23a5e313da0755ca285077a74 to main.

Imported generic function and interface/class method signatures now retain local type parameters, constraints, and defaults across source-owned program schemas. A focused cross-module control now matches pinned tsc and tsgo exactly (only the deliberate TS2322; no false TS7006).

On the exact pinned Zod 4.5.2 corpus, deterministic diagnostics improved from 1,349 to 1,347 (two removed, zero added). Zod is not yet semantically admitted, so timing remains intentionally blocked and the README benchmark table was not given a misleading result.

23 days ago

Progress via #548: pushed https://github.com/home-lang/home/commit/18f936ab6237818800fd65c6470f1749f9ffcef0 to main.

Primitive keyword type arguments now remain intrinsic during unresolved-generic recovery even when same-spelling runtime functions exist. A focused reference-compiler control retains a genuine value-only TS2749 negative control.

On the exact pinned Zod 4.5.2 graph, a fresh same-parent ReleaseFast A/B is deterministic across two runs each: 1,348 diagnostics on parent 3852cd38f versus 1,323 on candidate 18f936ab6. The exact diff removes 25 false primitive TS2749 diagnostics and adds zero.

This also corrects the prior pre-commit snapshot count: the exact rebuilt 3852cd38f source has one additional TS7006 at classic/schemas.ts:1631, so 1,348 is the authoritative parent count. That contextual-typing gap remains visible. Zod is still not timing-eligible, so no README timing row was added.

23 days ago

Zod parity update: pushed https://github.com/home-lang/home/commit/7ef4a49a070d75621f88094858b6e0a440dc0260 to main, fixing contextual typing through named namespace re-exports whose resolver already supplied a canonical owner path. The untouched 106-file Zod 4.5.2 corpus is deterministic at 1,322 diagnostics, down from the verified 1,323 checker baseline; pinned tsc 6.0.3 and tsgo 7.0.2 remain silent. Full evidence and validation: /home-lang/home/issues/548#issuecomment-5520257374. No timing or README benchmark row yet because semantic parity is still incomplete.

23 days ago

Zod admission progress: c71870a resolves #588 and removes 156 false TS7006 diagnostics with zero additions on the untouched pinned 106-file corpus (1,322 -> 1,166, deterministic output hash 6f30c07c30b804eeb88181108347bfd5e11083d9000467fc6157906d62365ad6). Pinned tsc 6.0.3 and tsgo 7.0.2 remain silent. The final nine nested-callback TS7006 diagnostics are tracked in #589; timing/README publication remains correctly blocked until semantic parity.

23 days ago

Benchmark-parity progress: 90a9f5a fixes complete Array callback arity under #589. The pinned unchanged Zod corpus moves 1,166 -> 1,164 diagnostics with 0 additions, deterministically. The benchmark remains unadmitted and no timing/README row is published while nine TS7006 diagnostics remain.

23 days ago

Zod parity progress: aa659e5 removes the two false JSON transform callback TS7006 diagnostics with zero additions (1,164 -> 1,162, deterministic hash fcfb4df6c0f270edc09a9a6287be10eb0b4c1690df1ce0d198fb63089b24bd4a). Seven #589 TS7006 diagnostics remain, so the benchmark is still not admitted and no timing/README row is published.

23 days ago

TypeScript correctness progress: a5ae69070 landed directly on origin/main and completes #589/#609. The admitted Zod 4.5.2 baseline is now 1,148 diagnostics (from 1,149), with exact 0-added/1-removed parity delta, deterministic repeated output, and no emit. This is a correctness admission only; no new timing claim or benchmark-table change is being made.

23 days ago

#540 is complete on main in b87db62c6, 1743ca880, and e53d7e845.

The export-list correctness gate was expanded from 96 to 128 cases to include renamed-local direct and two-hop barrel controls. TS 6.0.3, native TS 7.0.2, and Home now each pass 128/128 (384/384 compiler checks). Pristine base Home passed 80/128. Direct owner misses now retain TS2459/TS2460 provenance, while barrel misses remain TS2305.

This is an untimed admission improvement, not a new performance result; the existing 20/20 macOS and Linux benchmark snapshot is unchanged. The detailed matrix/evidence is linked from the new README admission table and recorded under the “Indexed export queries and variable-list ownership” section of TS_PERFORMANCE.md.

23 days ago

Main advanced through #617: returned imported callable factories now retain contextual callback types. On the exact Zod 4.5.2 graph this removes 60 TS7006 diagnostics with zero additions (1,148 → 1,088); all other diagnostic counts are unchanged. Two ReleaseFast outputs per arm are deterministic, the focused Program suite is 172/172, and README/detail docs are updated in 77deef9. #548 remains open and untimed while the remaining 1,088 diagnostics are reduced through separately controlled child issues.

23 days ago

Opened #619 from the next post-#617 Zod cluster: imported homomorphic mapped arguments lose contextual method types when the explicit local type argument contains a polymorphic this member. The reference-controlled reduction and required no-reparse, shadowing, negative-code, and exact Zod A/B gates are in the issue.

23 days ago

Completed child slice #619 and shipped it to origin/main through ad4ae62.

The change adds exact mapped-receiver controls shared by TypeScript 6.0.3, native TypeScript 7.0.2, and Home, plus a documented pinned-Zod same-correctness audit: 112 diagnostics removed, zero added, and 42.2% lower mean wall time versus exact parent 77deef9. This does not admit Zod into the cross-compiler speed table; #548 remains open until Home reaches the zero-diagnostic oracle.

23 days ago

Next evidence-led correctness child is #624 under #548. Post-#619, TS7006 is the largest remaining pinned-Zod class at 368. A strict imported/barrel reduction isolates loss of optional, generic, and rest parameter tuples through conditional mapped prototype members; TypeScript 6/7 agree on only the deliberate TS2339, while Home adds eight implicit-any diagnostics. Timing remains gated behind zero-added semantic evidence.

23 days ago

#624 is complete and landed on main as b73c3a20a..4deb785cc.

On the pinned 106-file Zod 4.5.2 graph, Home moved from 976 to 833 diagnostics: 143 removed (127x TS7006, 16x TS7019), with zero normalized additions. Program 174/174, Checker 4,334/4,334, ReleaseFast, formatting, scoped lint, and the final byte-identical integration audit all pass. The README links to the full benchmark record.

Zod still has 833 diagnostics, so the broader zero-diagnostic admission work remains open.

Sign in to comment on this issue.