From bebdda7850cc0771cb17623ccf388788564926b8 Mon Sep 17 00:00:00 2001 From: Gergo Magyar Date: Tue, 11 Aug 2026 18:30:56 +0000 Subject: [PATCH] perf(kotlin): intern dirChildren keys per directory and compact the buckets MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Two optimizations to `getKotlinFileIndex`, both output-identical, kept because they were measured and a third was dropped because it was not. 1. PER-DIRECTORY KEY MEMO. The component walk over `dir` cut one `slice` per component per FILE, and every slice after the first file of a directory is a freshly allocated string that hashes to a key the map already holds and is then dropped. The key list is a pure function of `dir`, so it is interned once per DIRECTORY. Measured -18.4% to -21.7% of the build at 32 000 files; zero retained cost, the memo dies with the frame. 2. BUCKET COMPACTION at the freeze loop. `addChild` mints `[raw]` and pushes, and V8 grows a backing store by `old + old/2 + 16`, so the SECOND child takes a 1-slot store to 17 and every bucket then retains its overshoot. 61 144 buckets at 32 000 files, 52.9% of their slots empty, 88 B each. `slice()` on freeze: -5 397 768 B, -11.20%, and the predicted 5 382 507 B lands within 0.03% of it. Same fix and the same accounting as the python `byBasename` note this repo already carries. `length === 1` is skipped deliberately. A bucket that never grew is already exact, so slicing it allocates a second array to save nothing — on a corpus of single-file packages the unguarded form costs 31% of the build for zero bytes. DROPPED: merging the `dirChildren` walk into the `suffixByStem` walk. It measures -0.10% at 32k, +0.23% at 100k and +0.40% at one file per directory, all inside a base-vs-identical-copy noise floor of -2.3% to +3.1%, and it does not compose usefully with the memo — the second scan it deletes is exactly the scan the memo makes rare. Only its provably free half is kept: `stem.lastIndexOf('/')` in place of `norm.lastIndexOf`, one backwards scan instead of two, exact because an extension carries no '/'. Neither optimization is visible to the correctness fingerprint, which is the point and also the risk: it observes the index only through the four resolver tiers, so a key-order move no corpus query reaches would survive it. Correctness therefore rests on a structural comparison of all three maps — key insertion order, values, bucket contents in order, frozen-ness — over 1234 corpora in both iteration orders, 14 808 comparisons, zero failures. The fingerprint, `cases` and `non_null` are unchanged and MUST NOT be re-baselined by this commit. Gates that did move, both because a reading and its budget move with the code rather than when CI goes red: - `heap_reading_bytes.kotlin` 48 200 224 -> 42 802 456 with its ceiling at 1.5x. A memory WIN passes every arm, so nothing forced this. - `depth_budget` 2.2 -> 2.0. The memo turns a per-file component walk into a per-directory one, which is precisely the per-depth work this arm exists to see: the band went 1.27-1.40 -> 1.20-1.26, and 2.2 held over it would have drifted from ~1.6x headroom to ~1.9x. The gate controls were re-run against the optimized builder, including one this change makes newly plausible: keying the memo on the directory's LAST SEGMENT instead of its full path drifts the fingerprint (36a4e9dad313, non_null 13310 -> 13305). That is the memo's whole safety argument stated as a test — its key decides which key set a directory contributes — and it is the one way this optimization could move an answer. The bucket-cap control was re-run too, since compaction now rewrites the same buckets. Also recorded, from measuring a reuse this repo had been invited to make: replacing `dirChildren` with the shared `package-dir-index` is output-identical (0 divergences over 107 948 answers) and passes every arm of the kotlin bench at 1.37x-1.50x — while costing 409x per fan-out and 8114x on `import data.*` at 200 matching directories on a corpus this bench does not carry. `_blind_spot` in the kotlin baselines now says so, with the memory the trade would have bought (26.2%, 12.18 MiB) and the corpus arm that would have to exist first. Refs #2881. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01W8QYxYvd5ntikRUvpLZD5e --- gitnexus/bench/import-target/baselines.json | 8 +-- gitnexus/bench/import-target/measure.mjs | 16 +++-- .../bench/kotlin-import-target/baselines.json | 15 ++-- .../languages/kotlin/import-target.ts | 71 ++++++++++++++++--- .../kotlin-import-target-parity.test.ts | 8 ++- 5 files changed, 91 insertions(+), 27 deletions(-) diff --git a/gitnexus/bench/import-target/baselines.json b/gitnexus/bench/import-target/baselines.json index b55ae8ccd..8dd560306 100644 --- a/gitnexus/bench/import-target/baselines.json +++ b/gitnexus/bench/import-target/baselines.json @@ -1,6 +1,6 @@ { "_what": "Baselines for bench/import-target/measure.mjs \u2014 EVERY import-target resolver registered in SCOPE_RESOLVERS, on one shared corpus, plus csharp a second time WITH csproj configs. One entry per registered language and one more for the csproj arm, no registered language ungated \u2014 and that is ASSERTED rather than asserted-in-a-comment, which is also why no roster of language names is kept in this prose to go stale: measure.mjs derives its language list from a LANG_REGISTRY table and a --check inventory arm reconciles that table against SCOPE_RESOLVERS in both directions. A C/C++ #include is an import site for this purpose and is gated like every other registered language. csharp and csharp_csproj resolve the IDENTICAL file corpus (buildFiles aliases the two) and differ in exactly one thing: whether csharpConfigs is supplied. Without that second arm the csproj namespace-directory index ships unmeasured, because every C# import in the no-csproj arm returns before reaching it. C and C++ follow that same precedent for a different context \u2014 their HEADERS arrive through resolutionConfig rather than through allFilePaths, and augmentedFilePaths unions the two once per pass, so the corpus is split at newPass rather than pre-merged. The first nine were added as their own O(imports x files) scans were indexed away (#2877/#2878/#2879/#2880, #2872, #2901, #2902, #2908) and this is the forward guard on each; the other eight were ungated until now, and PR #2911 \u2014 JavaScript reaching suffixResolve with no index at all, 25972 us per import at 8000 files \u2014 is what that costs.", - "_fingerprint_note": "Per-language sha256 over every distinct fromFile|target -> resolved target. A change here is a BEHAVIOUR change: the resolver returned a different target set, and IMPORTS/CALLS edges moved. Explain it, never re-baseline to make CI green. For the languages these PRs changed, the pre-change implementations produce these same values on this corpus at both 400 and 1600 files \u2014 that is what makes the index hoist a performance change. The tie-break-level proof lives in test/unit/scope-resolution/import-target-index-parity.test.ts (verbatim copies of the pre-change code, diffed) for Kotlin in test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts, and for the four resolvers added there in test/unit/scope-resolution/{php,java,cobol}-import-target-parity.test.ts and test/unit/import-resolvers/csharp-csproj-parity.test.ts, and for JavaScript in test/unit/scope-resolution/javascript-import-target-parity.test.ts (a differential over 211200 old-vs-new pairs, PR #2911). The eight languages added last have no per-language parity harness against a pre-change implementation and do NOT need one: nothing about their resolution changed, so there is no before to diff against. Their fingerprints are pure forward guards, minted from the current implementations, and their adapter-boundary index reuse is covered for every registered language at once by test/unit/scope-resolution/import-target-index-reuse.contract.test.ts. NOTE for csharp_csproj: on this corpus the #2902 indexed leg (step 3 of resolveCSharpImportInternal) is reached by 2221 of the 3200 small-arm imports but answers null for every one of them \u2014 the 979 that resolve do so at step 2 \u2014 so this fingerprint pins that legs cost and its null answers, while its positive tie-breaks (unanchored substring, iteration order) are pinned by csharp-csproj-parity.test.ts. NOTE for kotlin, go, csharp and java: twenty fingerprints across these four languages were re-baselined in #2881, the one deliberate behaviour change any language in this file has had. It landed in two steps and the second is the reason the first is not a special case: Kotlin first, then the shared package-dir-index (go, java, csharp) and the csproj namespace index once the same rule was found live there. `getKotlinFileIndex` no longer requires a file's package directory to be the FIRST occurrence of that name in its own path, so the unique arm's `d % 7` nested slice (`mod{d}/src/main/kotlin/com/example/pkg{d}/inner/pkg{d}`) now belongs to package `pkg{d}` and its wildcard imports resolve: resolved 1100 -> 1153 small and deep, 4456 -> 4681 large. The collide arm needed a CORPUS edit alongside it, not just a new number \u2014 its `d % 7` slice deliberately imported `com.example.vendor{d}`, a package that exists nowhere, purely to mirror the unique arm's nested-slice MISS, so leaving it would have left collide at 1100 against small's 1153 and broken the same-workload invariant the arm is built on (that assertion is what caught it). It now uses the same `com.example.models.*` spelling as the rest of the arm, which is why its distinct_outcomes fell (2775 -> 2744, 11087 -> 10961): one shared target instead of one per d. The record-level evidence for the resolver change \u2014 235 of 19968 records moved, 54 null -> resolved, 0 buckets losing a member \u2014 is in bench/kotlin-import-target/baselines.json `_provenance`. The kotlin heap_reading_bytes and heap_ceiling_bytes moved with it, together as `_heap_reading_note` requires: 48073096 -> 48200224 bytes_large (+127128, +0.264%), ceiling still exactly 1.5x. Small, and it is worth saying WHY it is small rather than reading the number as evidence that the change is cheap. `dirChildren` grows by one entry per component-suffix the old rule used to skip, and this arm can only see part of that: the heap corpus is built with HEAP_PAD 8, which prefixes every path with `d0/\u2026/d7/`, so no path can begin with a suffix of its own directory and the leading-segment half of the old rule is structurally invisible here. What moves the reading is the `d % 7` nested slice alone. Read +0.264% as this arm's ceiling on the effect, not as the effect. GO NEEDED A CORPUS EDIT TO BE GATED AT ALL. Its nested slice was `src/pkg{d}/internal/pkg{d}`, repeating only the LAST segment, while a Go query addresses the whole package path `src/pkg{d}` \u2014 so the directory never even ended with the query and the first-occurrence rule was never reached. Every go arm sat unchanged through the resolver fix. `uniqueDir`/`collideDir` now repeat the shape at the granularity Go actually queries (`src/pkg{d}/internal/src/pkg{d}`, `svc{d}/internal/sub/svc{d}/internal`), which is what moved go from 979 to 1153 resolved and bumped `languages.go.heap.path_segments` 13 -> 14. The general lesson: a corpus that carries a shape the QUERY cannot express does not gate that shape. CSHARP AND JAVA HIT THE SAME COLLIDE-ARM TRAP AS KOTLIN. Both collide arms sent their `d % 7` slice to a namespace that exists nowhere (`App.Src{d}.Vendor`, `com.svc{d}.vendor`) purely to MIRROR the unique arm's nested-slice miss; once that miss became a hit, collide sat at 979/1100 against small's 1153 and the same-workload assertion failed. Both now use the same spelling as the rest of their arm. HEAP: only `csharp_csproj` moved measurably (73703384 -> 73116520, -0.79%, stable across runs), re-recorded here with its ceiling; the step-2 filter change alters which lazily-built `getFilesInDir` maps the probe import forces, and that laziness is what the reading sees. Everything else sat within +/-0.03%. Note that `_heap_reading_note`'s claim that these readings 'reproduce to the byte across processes on one box' did NOT hold on the box this was measured on: go, dart, ruby, python, php and cpp all wandered by a few hundred to a few thousand bytes between processes with no code change touching them. Treat sub-0.05% movement as jitter, not signal.", + "_fingerprint_note": "Per-language sha256 over every distinct fromFile|target -> resolved target. A change here is a BEHAVIOUR change: the resolver returned a different target set, and IMPORTS/CALLS edges moved. Explain it, never re-baseline to make CI green. For the languages these PRs changed, the pre-change implementations produce these same values on this corpus at both 400 and 1600 files \u2014 that is what makes the index hoist a performance change. The tie-break-level proof lives in test/unit/scope-resolution/import-target-index-parity.test.ts (verbatim copies of the pre-change code, diffed) for Kotlin in test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts, and for the four resolvers added there in test/unit/scope-resolution/{php,java,cobol}-import-target-parity.test.ts and test/unit/import-resolvers/csharp-csproj-parity.test.ts, and for JavaScript in test/unit/scope-resolution/javascript-import-target-parity.test.ts (a differential over 211200 old-vs-new pairs, PR #2911). The eight languages added last have no per-language parity harness against a pre-change implementation and do NOT need one: nothing about their resolution changed, so there is no before to diff against. Their fingerprints are pure forward guards, minted from the current implementations, and their adapter-boundary index reuse is covered for every registered language at once by test/unit/scope-resolution/import-target-index-reuse.contract.test.ts. NOTE for csharp_csproj: on this corpus the #2902 indexed leg (step 3 of resolveCSharpImportInternal) is reached by 2221 of the 3200 small-arm imports but answers null for every one of them \u2014 the 979 that resolve do so at step 2 \u2014 so this fingerprint pins that legs cost and its null answers, while its positive tie-breaks (unanchored substring, iteration order) are pinned by csharp-csproj-parity.test.ts. NOTE for kotlin, go, csharp and java: twenty fingerprints across these four languages were re-baselined in #2881, the one deliberate behaviour change any language in this file has had. It landed in two steps and the second is the reason the first is not a special case: Kotlin first, then the shared package-dir-index (go, java, csharp) and the csproj namespace index once the same rule was found live there. `getKotlinFileIndex` no longer requires a file's package directory to be the FIRST occurrence of that name in its own path, so the unique arm's `d % 7` nested slice (`mod{d}/src/main/kotlin/com/example/pkg{d}/inner/pkg{d}`) now belongs to package `pkg{d}` and its wildcard imports resolve: resolved 1100 -> 1153 small and deep, 4456 -> 4681 large. The collide arm needed a CORPUS edit alongside it, not just a new number \u2014 its `d % 7` slice deliberately imported `com.example.vendor{d}`, a package that exists nowhere, purely to mirror the unique arm's nested-slice MISS, so leaving it would have left collide at 1100 against small's 1153 and broken the same-workload invariant the arm is built on (that assertion is what caught it). It now uses the same `com.example.models.*` spelling as the rest of the arm, which is why its distinct_outcomes fell (2775 -> 2744, 11087 -> 10961): one shared target instead of one per d. The record-level evidence for the resolver change \u2014 235 of 19968 records moved, 54 null -> resolved, 0 buckets losing a member \u2014 is in bench/kotlin-import-target/baselines.json `_provenance`. The kotlin heap_reading_bytes and heap_ceiling_bytes moved with it, together as `_heap_reading_note` requires: 48073096 -> 48200224 bytes_large (+127128, +0.264%), ceiling still exactly 1.5x. Small, and it is worth saying WHY it is small rather than reading the number as evidence that the change is cheap. `dirChildren` grows by one entry per component-suffix the old rule used to skip, and this arm can only see part of that: the heap corpus is built with HEAP_PAD 8, which prefixes every path with `d0/\u2026/d7/`, so no path can begin with a suffix of its own directory and the leading-segment half of the old rule is structurally invisible here. What moves the reading is the `d % 7` nested slice alone. Read +0.264% as this arm's ceiling on the effect, not as the effect. GO NEEDED A CORPUS EDIT TO BE GATED AT ALL. Its nested slice was `src/pkg{d}/internal/pkg{d}`, repeating only the LAST segment, while a Go query addresses the whole package path `src/pkg{d}` \u2014 so the directory never even ended with the query and the first-occurrence rule was never reached. Every go arm sat unchanged through the resolver fix. `uniqueDir`/`collideDir` now repeat the shape at the granularity Go actually queries (`src/pkg{d}/internal/src/pkg{d}`, `svc{d}/internal/sub/svc{d}/internal`), which is what moved go from 979 to 1153 resolved and bumped `languages.go.heap.path_segments` 13 -> 14. The general lesson: a corpus that carries a shape the QUERY cannot express does not gate that shape. CSHARP AND JAVA HIT THE SAME COLLIDE-ARM TRAP AS KOTLIN. Both collide arms sent their `d % 7` slice to a namespace that exists nowhere (`App.Src{d}.Vendor`, `com.svc{d}.vendor`) purely to MIRROR the unique arm's nested-slice miss; once that miss became a hit, collide sat at 979/1100 against small's 1153 and the same-workload assertion failed. Both now use the same spelling as the rest of their arm. HEAP: only `csharp_csproj` moved measurably (73703384 -> 73116520, -0.79%, stable across runs), re-recorded here with its ceiling; the step-2 filter change alters which lazily-built `getFilesInDir` maps the probe import forces, and that laziness is what the reading sees. Everything else sat within +/-0.03%. Note that `_heap_reading_note`'s claim that these readings 'reproduce to the byte across processes on one box' did NOT hold on the box this was measured on: go, dart, ruby, python, php and cpp all wandered by a few hundred to a few thousand bytes between processes with no code change touching them. Treat sub-0.05% movement as jitter, not signal. HEAP, kotlin, second movement: 48200224 -> 42802456 (-11.20%), re-recorded with its ceiling. `getKotlinFileIndex` now compacts each `dirChildren` bucket as it freezes it. `addChild` mints a bucket as `[raw]` and pushes the rest, and V8 grows a backing store by `old + old/2 + 16`, so the second child takes a 1-slot store to 17: 61144 buckets, 52.9% of their slots empty, 88 B each. Same fix and same accounting as the python `byBasename` sentence above. Note what this means for the gate: a memory WIN of this size passes every arm \u2014 it is under the ceiling and over the 0.5x floor \u2014 so it is recorded because the convention says a reading and its ceiling move together, not because anything went red. kotlin now reads 40.82 MiB. The prose in measure.mjs calling it '45.85 MiB, the second-largest reading in this file' is corrected with it \u2014 and was already wrong on the ranking before this change, since csharp_csproj (69.73) and php (47.28) both read higher; kotlin was third. A measurement written into prose is not re-taken, which is the finding `_heap_bound_note` records about this very file.", "_shape_note": "files/imports/resolved/distinct_outcomes AND the fingerprint are asserted exactly, per scale. A fingerprint alone cannot tell a legitimate resolution change from a corpus quietly shrunk below the size at which the timing arms can see anything; conversely the counts alone cannot see a defect confined to one arm, because the arms differ only in path padding and directory layout and both of those are count-neutral by design. Two cross-arm assertions close the remaining hole: the deep and collide arms must resolve exactly what small resolves (they are the same workload), and each of their fingerprints must DIFFER from small's (they are not the same corpus). Without the second, setting DEEP_PAD to 0 \u2014 which deletes the entire depth arm \u2014 moves no asserted number and prints PASS; the same is true of a collideDir that forwards to uniqueDir. THE HEAP ARM IS ASSERTED THE SAME WAY, by the same loop, and was not before: files_small, files_large, path_segments and probe decide WHAT it measures, and every one of them was reported and compared to nothing. Swapping HEAP_PROBE_TARGET.csharp_csproj for a target matching no CSPROJ_CONFIGS rootNamespace skips the whole config loop, so the getFilesInDir and getInsensitive legs never run and the arm the header calls the witness that the read pattern IS the footprint quietly becomes a two-map arm \u2014 73703384 -> 59921216 B, ratio 1.017 -> 1.011, ceiling and floor both still passing and --check still exiting 0. Setting HEAP_SMALL equal to HEAP_LARGE is the same hole from the other side: ratio goes to ~1.0 by construction and bytes_large never moves. bytes_small and bytes_large are deliberately NOT asserted for equality \u2014 heap_ceiling_bytes and the heap_reading_bytes floor bound them with ~50% either way, because heapUsed accounting moves across platforms and Node majors and an exact byte assertion would be a re-baseline per runner. THE CONTEXT ARM IS ASSERTED THE SAME WAY, by the same loop, and more strictly than either: target, with_context and without_context are exact strings with no tolerance at all, because the arm resolves one import over a three-file corpus and has no measurement noise to tolerate. A separate check requires the last two to DIFFER, for the same reason deep.fingerprint must differ from small.fingerprint \u2014 a probe on which both call shapes agree asserts one number twice. Both halves run through resolveOne, so what the arm gates is this bench threading run.ts's fifth argument, not the resolvers' behaviour.", "_arms_note": "Five timing arms, one memory arm and one deterministic arm elsewhere, because none of them gates alone. scaling_ratio (t_large/t_small)/(1600/400) catches cost growing with FILE COUNT \u2014 the #2877-#2880, #2901, #2902 and #2908 regressions themselves; every one of those legs was Theta(files) per import, so a revert scores ~4 here by construction. depth_ratio (t_deep/t_small at a FIXED file count, ~6x the path components) catches cost growing with path DEPTH, which scaling_ratio divides out and structurally cannot see; buildSuffixIndex (C#, Ruby, PHP, Java) and Kotlin suffixByStem emit one entry per component, so they legitimately sit above 1.0 while Go, Dart and COBOL, whose indexes are depth-free, sit at ~1.0. csharp's depth_budget has now been retightened twice for the same reason, and the second time it did lock the win in. It was 5 against a then-measured 3.318; #2903 made buildSuffixIndex's dirMap lazy and it became 3.5 against 2.31, with the file stating plainly that 3.5 did NOT lock that win in because a revert to an eager dirMap scores 3.318 and passes. Extending the laziness to the two SUFFIX maps drops it again, to 1.438 (java likewise 2.214 -> 1.402), because the deep arm has ~6x the path components and an O(files x depth) build of a map the no-csproj leg never reads is exactly the cost that scales with depth. Both are now 2.2, which is this file's 1.5x convention against measurements whose own peak-to-peak over 4 runs is 1.04x and 1.07x \u2014 and 2.2 DOES lock it in: an eager rebuild scores 2.3+ and fails. The other fifteen depth budgets sit at 1.37-1.75x measured and are unchanged. collide_scaling_ratio is the same measurement on a SHARED-LEAF layout (svcN/internal, SrcN/Models, com/example/model in every service, a repeated mod0.dart/mod0.rb/Mod0.cpy basename) carrying an identical file, import and resolved count: the small/large/deep arms mint one directory name per index, so every index bucket in them holds exactly ONE entry (measured: max last-segment bucket 1 and max matching directories 1 for go and csharp at 400 and 1600 files; max basename bucket 1 for dart and ruby), and bucket cardinality is the only non-constant term the new indexes have. On the shared-leaf shape go, csharp, dart and java legitimately score 2.1-3.9 because the bucket grows with the file count BY CONSTRUCTION \u2014 this is a limit on the SCOPE of the \"independent of corpus size\" claim, not a regression (the indexed code is still faster there than the pre-change full scan); their collide budgets say so honestly instead of pretending 1.8. Ruby, Kotlin, PHP and COBOL answer from keyed maps and are collision-immune, so they keep the linear 1.8 budget and that immunity is the assertion. csharp_csproj is the one arm that runs the other way: its shared leaf collapses dirsByLastSegment to the single key Models, so the slash-free sweep (see CSPROJ_CONFIGS) is CHEAPER on the collide layout than on the unique one and its expensive scale arm is large, not collide_large. Its 1.8 collide budget is therefore the linear one, and the arm that carries its real cost is the unique one. The collide arm is also the only arm that reaches filesDirectlyInPkgDir's dirCount > 1 merge (go: 388 multi-directory calls at 400 files, up to 9 directories; 1517 at 1600 files, up to 34) and the only one that reaches COBOL's copybook-over-source tier tie-break, which needs one bookname to name two files. small_ms_ceiling and collide_ms_ceiling are ABSOLUTE (~4x the measured arm), because a constant-factor regression that grows both scale arms equally passes every ratio. The five arms added here use 4.2x, the middle of the 3.7-4.6x the original five already carry; the two COBOL arms use ~5x, the multiplier dart's sub-1 ms arm has always carried, because a fixed scheduler hiccup is a larger fraction of a smaller number \u2014 measured over 8 runs they sat at 0.25-0.37 ms and 0.18-0.30 ms, and the pre-#2908 two-scans-per-COPY implementation costs ~300 ms on the same arm, so 2.0 and 1.5 still separate fixed from broken by two orders of magnitude. NOISE, measured rather than assumed: depth_ratio divides two sub-3 ms numbers (Dart's are sub-1 ms) and is by far the noisiest arm here, so it set N for the whole file. fastest() is a min-of-N estimator, so N is the knob. Over 22 --check runs on an idle box, peak-to-peak: at N=5 go ran 0.757-1.748 (2.31x) and tripped its own 1.6 budget about 1 run in 20; at N=7 (the kotlin-import-target setting) Dart still ran 0.678-2.043 (3.01x) and tripped once; at N=15 (bench/cfg, bench/schema-pairs, bench/callable-value-flow) every language collapsed to a 1.13-1.26x swing with 22/22 passing. The budgets were NOT widened; the estimator was fixed instead, which is why the headroom above is real rather than granted. N IS NOW PER LANGUAGE, and that is a refinement of the same finding rather than a retreat from it. The overshoot of min-of-K against min-of-15 is a function of the CELL's absolute duration, not of the language: replayed against two independent runs' full sample sets, the worst overshoots at K=7 land on swift.small (0.43 ms, 31.8%) and dart.collide (1.5 ms, 37.6%), while every cell at or above 10 ms overshoots by at most 6.3%. So repsFor() keeps 15 while a language's cheapest arm is under 5 ms and otherwise spends ~150 ms per cell, floored at 7 \u2014 15 for go, csharp, dart, kotlin, java, cobol, swift, rust, python, c and cpp (every language the flakiness above was ever about, cheapest arm 0.19-3.2 ms) and 7-8 for csharp_csproj, ruby, php, javascript, typescript and vue (cheapest arm 20-28 ms). Per LANGUAGE, not per cell, so all five arms of a language share one estimator and the four ratios stay comparisons of like with like. The replay passed all 85 cells on all five gates at 0.4-0.7 of budget and saved 12.8 s and 12.4 s of a 46 s run; min-of-7 also reads slightly HIGHER than min-of-15, so the ceilings get marginally more sensitive rather than less. Confirmed on 4 fresh runs with the adaptive estimator live: every small arm inside 1.12x peak-to-peak and every collide arm inside 1.07x, with the six 7-8 rep languages at 1.008-1.071 \u2014 no worse than the 11 that kept 15. The chosen N is reported per language as `reps`. heap_ceiling_bytes bounds the retained per-pass import index, the only arm here that can see memory: buildSuffixIndex emits maps at O(files x depth), the profile package-dir-index.ts cites #2649 to avoid for itself, and csharp, ruby, php and java all retained NOTHING across imports at BASE (C#'s no-csproj leg and PHP's and Java's every leg re-scanned the raw Set; Ruby rebuilt and discarded a suffix index per require). It is measured at 8000 and 32000 files at HEAP_PAD depth rather than at the timing arms' sizes, because the finding is an ABSOLUTE footprint at repository scale. THE ARM NOW READS WHAT THE LANGUAGE READS, and that change is the whole reason this file was re-baselined. Four of these arms used to call getWorkspaceFileIndex(set) directly and then read index.all.length, which asks no suffix question at all \u2014 harmless only while buildSuffixIndex built both maps eagerly. The moment they went lazy the direct call built NO map, csharp, ruby, php and java each reported 0 B at 32000 files, and 0 B is under every ceiling: --check printed PASS over four gates that had silently become ceilings over nothing, which is precisely the failure this file's own header warns about for rust and cobol. Every arm now resolves a real MISSING import through the real resolver (HEAP_PROBE_TARGET, asserted to miss), so the maps it forces are the maps production forces, and a resolver that starts asking a new question moves the number without anyone editing the bench. That makes the READ PATTERN the dominant term, and the eight numbers say so: java 34958600 B and csharp 29862200 B ask index.get and never getInsensitive; php 37579888 B asks getInsensitive and never get, plus its own first-proper-suffix map; ruby 41025360 B and javascript 26745296 B read get(s) || getInsensitive(s) and pay for both, the second DERIVED from the first; and csharp_csproj 73705944 B additionally asks getFilesInDir. csharp_csproj IS NOW GATED, reversing the earlier decision that it would be 'a ceiling on a duplicate': at +20.8% of the C# index it was one, and at 2.47x of it \u2014 same corpus, same getWorkspaceFileIndex, three maps instead of one \u2014 it is the witness that the read pattern is the footprint. The old RESIDUAL note is superseded by that number: a dirMap-sized addition is no longer +18%, and a consumer that asks all three questions blows csharp's ceiling by 1.64x rather than sliding under it. A SECOND MEASUREMENT BIAS was removed at the same time and it moved every figure here, so do not read these against the old ones as if only the read pattern changed. buildFiles mints paths with template literals, which V8 keeps as ropes; the first traversal that slices one flattens it, allocating the flat string and dropping the rope's pieces, so a build measured over an unflattened corpus reports the index MINUS that net release \u2014 11% low, uniformly. bytes_small was read over a corpus a discarded warm-up pass had already flattened and bytes_large over a fresh one, so every ratio read ~0.85-0.89 for structures that are exactly linear in the file count. measureHeap now flattens each corpus before measuring it; all eight ratios read 0.998-1.017, and the warm-up pass is gone because with the corpus flat a language's first and second reads agree to within 0.3%. python's figure rises from 7624992 to 10362976 for this reason and not because anything regressed, and then to 10543152 (+1.7%) because #2913's nestedDirNames set is retained for the pass, and then FALLS to 6360936 (-39.7%) for a reason worth knowing: byBasename holds roughly one bucket per file, and building each with `[]` followed by `push` made V8 grow the backing store to its 16-slot minimum, so every single-file bucket retained 15 empty pointer slots. Constructing the one-element buckets directly (`set(base, [entry])`) is byte-identical in contents and 3.9 MiB smaller at 32000 paths \u2014 37% of what this arm used to read was empty array slots \u2014 the ancestorsByDir memo itself is NOT in this reading, because python's probe target misses at the nested-name rejection and never reaches the walk, so this arm does not bound that memo; measured separately with a probe that does reach it, a 32000-file corpus with every file in its own 10-deep directory retains ~19 MB, which would clear this ceiling, so repointing python's heap probe at a walking spelling means re-recording the ceiling in the same change, and c is unchanged at 10018816 because its basename map does not slice paths. Its ceiling is 1.5x the measured arm, and the DIFFERENCE FROM THE 4x TIMING CONVENTION IS DELIBERATE \u2014 do not harmonise it back. 4x exists because runner contention dominates a wall-clock number; this one has essentially no measurement noise (across 4 runs the widest spread was 0.11% on python, 0.03% on csharp_csproj and 0.00% \u2014 identical to the byte \u2014 on ruby, php, java, javascript and c, and the same holds across separate processes), so 4x would throw away almost all of the gate's power and sail straight past the regression this arm exists to catch. 1.5x still tolerates ~50% of cross-platform and Node-version drift, far more than a Node major bump plausibly moves heapUsed accounting; it catches a duplicated index (+100%) or a second exactMap-sized suffix map (+~85%). heap_floor_fraction is the arm the 0 B incident proved was missing. A ceiling can only say 'not too big'; nothing said 'still measuring something', which is why four dead arms passed. The floor is 0.5 x each language's RECORDED READING (heap_reading_bytes), which is half the measured size and says so. It used to be 0.33 x the CEILING, described the same way \u2014 true only while every ceiling stayed at exactly 1.5x its reading, a convention this file states and nothing enforces, so re-tuning one ceiling upward would have loosened that language's floor by the same factor in the one direction a floor exists to watch. The two forms agree to within 0.8% for all eight today, so this is a correction of derivation, not of strength. It sits ~400x above the readings' own reproducibility and far below any collapse. A genuine 2x memory WIN trips it too, and that is intended: like a fingerprint move, it must be explained and re-baselined rather than absorbed. COBOL is left out for the opposite reason: its index is two Map, O(files) with no depth term, and at 32000 files its retained delta does not clear the noise of the measurement itself. heap_ratio_budget, the linear-growth check across the 4x file-count gap, is the orthogonal arm: it sees per-file and per-depth growth but not a constant factor. ---- THE EIGHT LANGUAGES ADDED LAST (swift, rust, python, javascript, typescript, vue, c, cpp) ---- They carry the SAME five arms and the same gates; what differs is which arm can actually fail for each, because each resolver has a different cost axis, and the budgets below say so instead of copying a number across. Every figure quoted is the MAXIMUM over 5 full runs on an idle box, and the peak-to-peak of every one of these arms stayed inside 1.10x over those runs \u2014 tighter than the 1.13-1.26x the original nine record, because none of these arms divides two sub-1 ms numbers the way dart depth_ratio does. depth_budget is ~1.5x measured throughout: swift 2.3 (1.487), rust 2.1 (1.377), javascript 2.1 (1.376), typescript 2.1 (1.381), vue 2.3 (1.563), c 3.0 (1.990), cpp 3.0 (1.999). PYTHON WAS 11 AGAINST 7.389 AND IS NOW 2.6 AGAINST 1.872, because #2913 fixed the resolver rather than the budget. Its INDEX was always depth-free; hasRepoCandidate and resolveAbsoluteFromFiles each rebuilt one ancestor prefix per directory component of the importer on EVERY import, and the index's own dirPrefixes build inserted one entry per component per file, so the resolver was quadratic in path depth where every other language here is linear or flat. The prefixes are a pure function of the importer's DIRECTORY, so they are now memoized per directory inside getPythonFileIndex (ancestorsByDir), the leading segment is rejected up front against a set of nested directory names, the module and package buckets are consulted before the walk rather than inside it, and the dirPrefixes build stops at the first ancestor already stored. All five fingerprints are byte-identical, so it is a hoist. The budget is 2.2, and BOTH numbers behind it were re-measured on a quiet box AFTER the context leg below started being measured, because that change moved the arm: the work it adds is depth-FLAT, so python's absolute cost more than doubled while depth_ratio FELL to 1.405-1.563 over 5 serial runs (peak-to-peak 1.11x). A budget carried over from before that change would have been slack against a smaller ratio. 2.2 is 1.41x the measured maximum, inside the 1.37-1.75x band the other fifteen sit in, and it LOCKS THE WIN IN: reverting the per-directory ancestor memo alone scores 2.524 and reverting the nested-name rejection alone scores 2.553, both measured under the current call shape, so each fails at 2.2 with 13% to spare. Do not read those two figures as the pre-#2913 cost \u2014 7.239 was that, and the gap closed because the bare-import tier stopped walking at all (see below). The other two parts of the fix are not gated by this arm and are not meant to be: reverting the bucket prune or the dirPrefixes early break lands under any budget this arm's noise supports, so they are gated deterministically instead, by the prefix-parity and package-probe arms of test/unit/scope-resolution/python/python-importer-ancestors.test.ts and python-import-target-parity.test.ts, which go red on exactly those two mutations. A timing budget catches what it can measure; the counts catch the rest. THE BARE-IMPORT TIER (`import os`, single segment, no dot) was a separate O(depth) walk in import-resolvers/python.ts that this bench cannot see at all, because every python arm here spells its imports with a dot and returns at the `pathLike.includes('/')` guard before reaching it. It ran TWICE per `from x import y` \u2014 the package probe's recursion re-ran the whole tail on identical inputs \u2014 and is now one memoized chain plus an O(1) proof-of-absence against the index's basename buckets: 12/24/72 Set probes at depth 1/4/16 became a flat 2, and 11.615 us/import at 18 path components became 0.740. Gated by probe COUNT in test/unit/scope-resolution/python/python-import-probe-count.test.ts, not here. collide_scaling_budget splits three ways. Three languages scan a bucket that grows with the corpus and get their measured value x1.5: swift 4.9 (3.279 \u2014 its bucket is the module file list it RETURNS, and its collide arm is four modules instead of dirs of them so that bucket is fileCount/4, i.e. 100 files at 400 and 400 at 1600), c 3.8 (2.535) and cpp 4.0 (2.639, the same basename bucket its suffix fallback walks). Four answer from keyed maps and keep the linear 1.8 \u2014 python 1.097, javascript 1.083, typescript 1.053, vue 1.079 \u2014 and that immunity IS the assertion, exactly as for ruby, kotlin, php and cobol. RUST IS THE ONE ARM THAT WAS REDESIGNED RATHER THAN BUDGETED. It resolves by probing candidate paths with allFilePaths.has(...) and never searches, so its cost is O(path segments) and provably flat in the file count (1.095 scaling, 1.061 collide scaling): a shared-leaf collide arm for rust would have asserted nothing, which is worse than no arm. Its collide corpus is instead a deep module tree (src/l0/l1/l2/l3/l4/mod{d}) whose targets carry ~2x the :: segments, so the arm exercises the axis that CAN grow, its 1.8 budget asserts the flatness across file counts, and collide_ms_ceiling 19 bounds the absolute cost of the long-path probe. small_ms_ceiling and collide_ms_ceiling are ~4x measured as everywhere else: rust 10/19 (2.609/4.704), python 7/8 (1.76/1.929, retightened from 12/15 against 3.044/3.771 by #2913), javascript 85/89 (21.254/22.145), typescript 85/86 (21.250/21.464), vue 81/93 (20.164/23.227), c 7/11 (1.620/2.850), cpp 7/12 (1.581/3.009). Swift takes ~5x (2 against 0.421 and 4 against 0.821) \u2014 the multiplier dart and cobol already carry, because a fixed scheduler hiccup is a larger fraction of a sub-1 ms number. ONE CAVEAT ON THE THREE ts-FAMILY MS NUMBERS, stated because nothing else in this file would reveal it: resolveTsTarget carries a per-pass resolveCache keyed currentFile::importPath, which no other resolver here has, and ~10% of this corpus is repeat pairs. Their us/import is therefore a slight underestimate of a cold resolve. It is left in rather than defeated because it is what the real pipeline does, and it is identical across all three so the arms stay comparable. HEAP for the eight: rust, swift, typescript, vue, cpp and cobol are still NOT gated, all of them measured before being left out. rust builds no index on this hook (16 B at 8000 files, 0 B at 32000); swift holds one pointer per file-times-segment and mints no strings, reading 0.98 MB at 8000 files against 0.29 MB at 32000 \u2014 a 4x larger corpus reading 3x SMALLER, which is what a measurement below its own noise floor looks like, and the same reading cobol gives (0.54 MB then 0 B); typescript and vue duplicate javascript through the same builder over the same-shaped corpus, and cpp duplicates c (10021320 against 10016960, 0.04% apart). Those four duplications are the ONLY exclusions that still rest on 'it would be a duplicate', and they are duplicates of a builder AND of a read pattern, which is the pairing csharp_csproj failed once the read pattern started to matter \u2014 if any of the four ever diverges in what it ASKS the index, it earns an arm the same way csharp_csproj just did. All eight gated arms are read the same way now (retainedPassBytes, one real import), so unlike before they are directly comparable to one another. WALL CLOCK \u2014 ~33-35 s in report mode, down from ~46 s, and ~44-45 s for --check, which is essentially UNCHANGED from ~46 s. Only report mode got faster; do not read the pair as 46 -> 42. The breakdown is worth having before anyone trims it. Timing arms: go 2.02, csharp 1.09, csharp_csproj 3.22, dart 0.41, ruby 2.90, kotlin 0.85, php 3.46, java 1.57, cobol 0.09, swift 0.46, rust 0.85, python 1.22, javascript 3.23, typescript 2.72, vue 2.89, c 0.86, cpp 0.91 (28.7 s, from 39.8 s: repsFor() accounts for all of it, and every second of it comes from the six languages whose cheapest cell is 20-28 ms); heap arms 3.43 s for SEVENTEEN languages, from 2.06 s for eight (every registered language is measured now; the nine added cost 1.37 s, of which kotlin alone is 0.57 s \u2014 see _heap_bound_note), and 2.1 s came from 3.0 s for seven when flattening retired the warm-up pass; module load 3.9 s. --check pays one import that report mode does not: the inventory arm loads pipeline/registry.ts, which drags in every registered scope resolver and its providers. Measured in isolation with the bench's own static imports already resident, that import costs 6.3-6.5 s on one box and 9.3-10.0 s on another \u2014 i.e. it consumes almost the whole repsFor win, which is why --check did not get faster. It is loaded dynamically at the point of use rather than at the top of the file, so report mode does not pay it and both modes take their measurements in the same module state. IT WAS WEIGHED AND KEPT, on the number that decides it: the benchmarks job is not CI's critical path. On the last green run of main it took 9 m 23 s against 12 m 58 s for the sharded coverage job that gates the merge, so ~4 m 40 s of slack sits above this bench and those seconds buy zero merge latency. Moving the arm to a vitest file would move the registry load ONTO the critical path, and would weaken it as well: this reconciles LANG_REGISTRY's SupportedLanguages values, which are what the five dispatcher branches key off, whereas a test that cannot import measure.mjs can only reconcile this file's arm NAMES plus a hand-written rule for de-aliasing csharp_csproj. The contract test import-target-index-reuse.contract.test.ts already covers the ADAPTER-boundary contract for every registered resolver; this arm covers a different claim, that the BENCH covers the pipeline. The ts family is still the largest single block of the timing phase (8.8 s) \u2014 its cost is suffixResolve probing ~39 extensions per path part on a miss, which is the real resolver and cannot be tuned away from the bench side. IF IT HAS TO SHRINK, drop collide and collide_large for typescript and vue and nothing else: -3.9 s, and it is the only cut that removes near-duplicate work rather than coverage, because all three run the same resolveTsTarget over the same buildSuffixIndex and javascript keeps the collide arm that covers their shared collision axis. Do NOT reach for REPS_MAX: it is 15 because depth_ratio tripped its own budget about 1 run in 20 at 5 and once at 7, and lowering it would re-open that for the eleven languages whose cheapest cell is sub-5 ms \u2014 which is where every recorded trip happened. The six languages it was safe to lower have already been lowered, per language and from a measurement, by repsFor(). ---- THE FIFTH ARGUMENT (context) AND THE TWO ARMS IT MOVED ---- resolveOne now makes run.ts's five-argument call for the two hooks that declare a fifth parameter, so php and python time the legs behind it. Nothing else moved: the other fifteen arms are handed no context and build no ParsedFile[] at all, and over five runs their five ms numbers and four ratios sit exactly where they did. Both languages' ten fingerprints, resolved counts and distinct_outcomes are IDENTICAL \u2014 the leg AGREES with the cascade on this corpus, which is the whole reason the context arm had to be added rather than leaving the fingerprint to notice. PHP: small_ms 27.762 -> 35.125 (+26.5%) and collide_ms 29.407 -> 36.182 (+23.0%), which is filesByDirectory plus, on every import that resolves, a candidate gather over the resolved file's directory and a localDefs filter; the ms ceilings keep PHP's own 4.21x and 4.26x multipliers (117 -> 148, 125 -> 154). depth_ratio 1.144 -> 1.283 and the 1.9 budget is UNCHANGED, which makes it 1.48x measured rather than 1.66x: directoryAliases emits one entry per path segment, so filesByDirectory is O(files x depth) and the depth arm is the only one that can see it \u2014 that budget got TIGHTER relative to its measurement, not looser, and 1.48x sits inside the 1.37-1.75x band the other sixteen carry. Its heap reading rises 37576816 -> 49574008 (+31.9%) for the same structure, and the reading is the MEMO rather than the workspace it indexes: newPass allocates the ParsedFile objects before retainedPassBytes takes its baseline sample, so they sit outside the delta. PYTHON, WHOSE FIGURES ARE THE LEAST SETTLED THING IN THIS FILE AND ARE RECORDED IN TWO SNAPSHOTS BECAUSE OF IT. A named import is the only spelling that reads context.parsedFiles, and it costs up to three entries into the resolver per import (package probe, exports check, submodule probe) where the synthetic namespace spelling this arm used to pass costs one. Against the resolver as it stood when the call shape changed that read small_ms 1.76 -> 5.751 and collide_ms 1.929 -> 5.894, ~3.1x. Against the resolver a few commits later \u2014 which stopped re-running the whole tail after a null package probe, a double-probe this bench could not previously see because the namespace spelling never entered that branch \u2014 the same arms read 4.404 and 4.505. The ceilings are 18 and 19, chosen to clear BOTH: 4.09x and 4.22x of the current numbers, 3.13x and 3.22x of the higher ones, so neither state is red. Retighten toward 4x once that resolver settles. ITS DEPTH ARM WAS DILUTED AND THE BUDGET IS RETIGHTENED TO MATCH, which is the one thing here worth arguing about: the added work is depth-FLAT, so depth_ratio FALLS 1.872 -> 1.478 while the absolute cost more than doubles, and 2.6 against 1.478 would be 1.76x \u2014 far looser than the 1.39x #2913 chose deliberately to lock its own fix in. 2.1 restores that multiplier (1.42x). THE TWO MUTATION SCORES #2913 RECORDED (3.123 for reverting the per-directory memo, 2.734 for reverting the nested-name rejection) WERE TAKEN AGAINST THE OLD CALL SHAPE AND HAVE NOT BEEN RE-TAKEN. Modelled forward, with the depth-quadratic term reappearing in every resolver entry so its absolute contribution scales with the entry count, they land near 2.8 and 2.4 \u2014 both above 2.1, and the second BELOW 2.6, which is the arithmetic that decided the budget. Re-run the two mutations before trusting the lock-in claim above. python's heap reading is unchanged (10543152 recorded; 10529848-10544616 across eight runs) because its probe misses before the branch that reads parsedFiles \u2014 see _blind_spot for why no probe can reach that memo. Every figure in this section is the MAXIMUM over its snapshot's runs (five, then three), with peak-to-peak 1.031-1.058 on php and 1.019-1.081 on python, taken on a box that was NOT idle and with another change landing in python's resolver mid-measurement. Re-take them serially before merging.", "_triage": "Every ratio and ms ceiling here is a TIMING signal \u2014 re-run on an idle machine before investigating; runner contention dominates. depth_ratio is the noisiest of them by a wide margin (it divides two sub-3 ms numbers, and Dart's are sub-1 ms): if exactly one arm fails and it is that one, suspect the machine first. N is 15 for every language whose cheapest arm is under 5 ms, rather than this bench's original 5, specifically to hold that arm's peak-to-peak swing under 1.26x \u2014 see _arms_note for the measured distributions and for why the six languages that drop to 7-8 are the ones where cell size makes it safe \u2014 so a depth_ratio failure that REPRODUCES is a real signal, not noise. Each language's chosen N is printed as `reps`; read it before blaming the estimator. The fingerprint, shape and heap arms are the opposite: deterministic (over 4 runs the heap arm's widest spread was 0.11% on python and 0.00% on java, javascript and c), a re-run never changes them, and they must never be wished away. TWO heap failures mean the arm STOPPED MEASURING rather than that memory grew, and both are deterministic: a heap floor failure says the probe no longer forces the index it used to (this is how four arms read 0 B when buildSuffixIndex went lazy, and 0 B passes every ceiling), and a `heap probe ... resolved` throw says a probe target that must MISS now hits, so the reading is a materialized answer and the legs past it were never reached. A heap BOUND failure is deterministic in the same way and means one specific thing: a language excluded from the budgeted tier has grown a structure, or started asking its index a question it did not ask when the exclusion was recorded \u2014 never a timing signal, never a re-run, and never fixed by raising the bound without saying what grew. The context arm is deterministic too, and a failure there means one specific thing rather than a range of them: run.ts's fifth argument is not reaching that resolver from this bench, or the leg behind it stopped running. Never a timing signal, never a re-run.", @@ -85,7 +85,7 @@ "heap_ceiling_bytes": { "vue": 43326024, "typescript": 40117944, - "kotlin": 72300336, + "kotlin": 64203684, "go": 4497696, "dart": 11751300, "cpp": 15035016, @@ -102,7 +102,7 @@ "heap_reading_bytes": { "vue": 28884016, "typescript": 26745296, - "kotlin": 48200224, + "kotlin": 42802456, "go": 2998464, "dart": 7834200, "cpp": 10023344, @@ -115,7 +115,7 @@ "javascript": 26745296, "c": 10018816 }, - "_heap_bound_note": "THE SECOND HEAP TIER. Every registered language is measured now; heap_bound_bytes gates the nine that are not BUDGETED above, and it gates them with one comparison and no floor. A ceiling says 'this index is not too big'. A bound says something narrower and it is the thing that was missing: 'the exclusion still holds' \u2014 this language has not grown an index since it was left out. measure.mjs's MEMORY section states the re-entry condition (if a language ever diverges in what it ASKS its index, it earns a budgeted arm) and until now nothing watched for the divergence; HEAP_LANGS was a hand-maintained list of eight whose two neighbours, LANG_REGISTRY and CONTEXT_LANGS, are both reconciled against a derived predicate in both directions. HEAP_BOUNDED is derived too \u2014 it is LANGS minus HEAP_BUDGETED \u2014 so the two tiers partition the languages and a new one cannot land outside both. WHAT RE-MEASURING FOUND, five runs each, maximum quoted, peak-to-peak in brackets. go 2998464 B [1.0021], dart 7834200 B [1.0006] and kotlin 48073096 B [1.0004] HAD NO STATED REASON AT ALL: the old prose opened 'SIX of the seventeen are deliberately NOT in HEAP_LANGS' against a list of eight of seventeen, and these three were the three nobody counted. All three retain a real per-pass structure (go's PackageDirIndex, dart's basename buckets, kotlin's suffixByStem cascade) and kotlin's 45.85 MiB is the second-largest reading in this file, above ruby's 39.12 and java's 33.34, both of which carry a full budget. swift 3449216 B [1.0024] and cobol 2320456 B [1.0000] were excluded as 'below the measurement's own noise floor' on readings of 0.29 MB and 0 B at 32000 files; they now read 3.29 MB and 2.21 MB, growing with the corpus (969120 B and 536264 B at 8000). Those old numbers were not wrong when taken \u2014 the ARM changed under them, when #2903's follow-up made every probe resolve a real import and when measureHeap began flattening its corpus \u2014 which is the whole finding: a measurement written into prose is not re-taken, and this file had already gone stale against itself, quoting javascript at 46208832 B four paragraphs after quoting it at 25.51 MiB. rust is the one exclusion that survived unchanged: 16 B at 8000 files and 16 B at 32000, identical in all five runs. typescript 26745296 B, vue 28884016 B and cpp 10023344 B are duplicates of a builder AND of a read pattern: typescript is byte-identical to javascript's 26745296 in four runs of five, cpp is +0.05% of c's 10018816, vue is +8.0% of javascript. HOW THE BOUNDS WERE CHOSEN. Eight of the nine take 1.5x their measured maximum, rounded up to the next 100000 B: go 4500000 (1.501x), dart 11800000 (1.506x), kotlin 72200000 (1.502x), cobol 3500000 (1.508x), swift 5200000 (1.508x), typescript 40200000 (1.503x), vue 43400000 (1.503x), cpp 15100000 (1.507x). 1.5x is NOT copied from the ceilings out of habit \u2014 it is the same number for a stated reason, and the reason is not noise: measured peak-to-peak on this box is at most 1.0024, so noise alone would justify 1.05x. What a bound has to survive is a RUNNER change, since heapUsed accounting moves across platforms and Node majors, and this file already fixes that allowance at 50% for exactly this measurement on exactly this arm. Using a second allowance for the same uncertainty on the same number would be two conventions, not more rigour. At 1.5x the bound catches what the re-entry condition is about \u2014 a language growing an index, which costs +85% for one more suffix map and +100% for a duplicate \u2014 and it does NOT catch a duplicate diverging by 8%. That limit is real and is stated rather than hidden: the tight form is a same-process ratio against the arm each duplicate is a duplicate OF, which is the only form immune to the drift the absolute bound has to tolerate. RUST TAKES AN ABSOLUTE BOUND INSTEAD, 1048576 B (1 MiB), because 1.5 x 16 B is 24 B and would fail on the first byte of anything \u2014 a multiplier on a reading that is already nothing is a gate that flakes rather than a gate that bites. 1 MiB is ~65000x the reading and still 2.2x below the smallest real index measured here (cobol's 2.32 MB at the same file count), so it separates 'builds nothing' from 'builds something' with room on both sides. NO FLOOR ON ANY OF THE NINE, and the reason differs by language rather than being uniform. For rust a floor would be a floor on noise. For the other eight the readings are stable enough to floor today, and for kotlin and dart \u2014 larger than budgeted arms \u2014 a floor would be worth having, since a lazily-built map going quiet is exactly how the four budgeted arms once read 0 B. Adding one is a PROMOTION to the budgeted tier, with a ceiling and a recorded reading beside it, not a line here: a floor whose companion ceiling does not exist asserts 'still measuring' against a number nothing else bounds. Recommended next, in order: kotlin, then dart, then go.", + "_heap_bound_note": "THE SECOND HEAP TIER. Every registered language is measured now; heap_bound_bytes gates the nine that are not BUDGETED above, and it gates them with one comparison and no floor. A ceiling says 'this index is not too big'. A bound says something narrower and it is the thing that was missing: 'the exclusion still holds' \u2014 this language has not grown an index since it was left out. measure.mjs's MEMORY section states the re-entry condition (if a language ever diverges in what it ASKS its index, it earns a budgeted arm) and until now nothing watched for the divergence; HEAP_LANGS was a hand-maintained list of eight whose two neighbours, LANG_REGISTRY and CONTEXT_LANGS, are both reconciled against a derived predicate in both directions. HEAP_BOUNDED is derived too \u2014 it is LANGS minus HEAP_BUDGETED \u2014 so the two tiers partition the languages and a new one cannot land outside both. WHAT RE-MEASURING FOUND, five runs each, maximum quoted, peak-to-peak in brackets. go 2998464 B [1.0021], dart 7834200 B [1.0006] and kotlin 42802456 B [1.0004] HAD NO STATED REASON AT ALL: the old prose opened 'SIX of the seventeen are deliberately NOT in HEAP_LANGS' against a list of eight of seventeen, and these three were the three nobody counted. All three retain a real per-pass structure (go's PackageDirIndex, dart's basename buckets, kotlin's suffixByStem cascade) and kotlin's 40.82 MiB is above ruby's 39.12 and java's 33.34, both of which carry a full budget. (It read 45.85 MiB when this was written, described here as 'the second-largest reading in this file' \u2014 it was third even then, behind csharp_csproj and php; #2881 later compacted its dirChildren buckets and took 11% off it. Same staleness this paragraph exists to document.) swift 3449216 B [1.0024] and cobol 2320456 B [1.0000] were excluded as 'below the measurement's own noise floor' on readings of 0.29 MB and 0 B at 32000 files; they now read 3.29 MB and 2.21 MB, growing with the corpus (969120 B and 536264 B at 8000). Those old numbers were not wrong when taken \u2014 the ARM changed under them, when #2903's follow-up made every probe resolve a real import and when measureHeap began flattening its corpus \u2014 which is the whole finding: a measurement written into prose is not re-taken, and this file had already gone stale against itself, quoting javascript at 46208832 B four paragraphs after quoting it at 25.51 MiB. rust is the one exclusion that survived unchanged: 16 B at 8000 files and 16 B at 32000, identical in all five runs. typescript 26745296 B, vue 28884016 B and cpp 10023344 B are duplicates of a builder AND of a read pattern: typescript is byte-identical to javascript's 26745296 in four runs of five, cpp is +0.05% of c's 10018816, vue is +8.0% of javascript. HOW THE BOUNDS WERE CHOSEN. Eight of the nine take 1.5x their measured maximum, rounded up to the next 100000 B: go 4500000 (1.501x), dart 11800000 (1.506x), kotlin 64300000 (1.502x), cobol 3500000 (1.508x), swift 5200000 (1.508x), typescript 40200000 (1.503x), vue 43400000 (1.503x), cpp 15100000 (1.507x). 1.5x is NOT copied from the ceilings out of habit \u2014 it is the same number for a stated reason, and the reason is not noise: measured peak-to-peak on this box is at most 1.0024, so noise alone would justify 1.05x. What a bound has to survive is a RUNNER change, since heapUsed accounting moves across platforms and Node majors, and this file already fixes that allowance at 50% for exactly this measurement on exactly this arm. Using a second allowance for the same uncertainty on the same number would be two conventions, not more rigour. At 1.5x the bound catches what the re-entry condition is about \u2014 a language growing an index, which costs +85% for one more suffix map and +100% for a duplicate \u2014 and it does NOT catch a duplicate diverging by 8%. That limit is real and is stated rather than hidden: the tight form is a same-process ratio against the arm each duplicate is a duplicate OF, which is the only form immune to the drift the absolute bound has to tolerate. RUST TAKES AN ABSOLUTE BOUND INSTEAD, 1048576 B (1 MiB), because 1.5 x 16 B is 24 B and would fail on the first byte of anything \u2014 a multiplier on a reading that is already nothing is a gate that flakes rather than a gate that bites. 1 MiB is ~65000x the reading and still 2.2x below the smallest real index measured here (cobol's 2.32 MB at the same file count), so it separates 'builds nothing' from 'builds something' with room on both sides. NO FLOOR ON ANY OF THE NINE, and the reason differs by language rather than being uniform. For rust a floor would be a floor on noise. For the other eight the readings are stable enough to floor today, and for kotlin and dart \u2014 larger than budgeted arms \u2014 a floor would be worth having, since a lazily-built map going quiet is exactly how the four budgeted arms once read 0 B. Adding one is a PROMOTION to the budgeted tier, with a ceiling and a recorded reading beside it, not a line here: a floor whose companion ceiling does not exist asserts 'still measuring' against a number nothing else bounds. Recommended next, in order: kotlin, then dart, then go.", "heap_bound_bytes": { "cobol": 3500000, "swift": 5200000, diff --git a/gitnexus/bench/import-target/measure.mjs b/gitnexus/bench/import-target/measure.mjs index 3722af112..e08516f41 100644 --- a/gitnexus/bench/import-target/measure.mjs +++ b/gitnexus/bench/import-target/measure.mjs @@ -270,9 +270,12 @@ * of eight, so go, dart and kotlin were excluded silently. All three * retain a real per-pass structure: go's `PackageDirIndex` reads * 2 998 464 B, dart's basename buckets 7 834 200 B, and kotlin's - * `suffixByStem` cascade 48 073 096 B (45.85 MiB) — the second-largest - * reading in this file, above ruby's 39.12 and java's 33.34, both of which - * carry a full budget. + * `suffixByStem` cascade 42 802 456 B (40.82 MiB) — above ruby's 39.12 and + * java's 33.34, both of which carry a full budget. (Read 48 073 096 B when + * this paragraph was written and described as "the second-largest reading + * in this file", which it was not even then: csharp_csproj and php both + * read higher. #2881 then compacted kotlin's `dirChildren` buckets and + * took 11% off it.) * 2. TWO OF THE STATED REASONS NO LONGER HOLD. swift was excluded as "below * its own noise floor" on 0.98 MB at 8000 files against 0.29 MB at 32 000; * it now reads 969 120 B and 3 449 216 B, growing the right way. COBOL was @@ -560,9 +563,8 @@ const HEAP_BUDGETED = [ // per-pass structure and each grows LINEARLY with the file count (ratio // 0.996-1.004 against a 1.25 budget over 8000 -> 32000 files), so each can // carry the full ceiling + floor + ratio set rather than a bound alone. - // kotlin's 45.85 MiB is the second-largest reading in this file — larger than - // ruby's and java's, both of which were budgeted from the start — and it had - // no stated exclusion reason at all. + // kotlin's 40.82 MiB is larger than ruby's and java's, both of which were + // budgeted from the start, and it had no stated exclusion reason at all. 'kotlin', 'dart', 'go', @@ -2608,7 +2610,7 @@ for (const lang of HEAP_BUDGETED) { * `1.5 x 0 B` is 0 — its bound is ABSOLUTE (1 MiB) for the same reason: a * multiplier on 16 B fails on the first byte of anything. The other eight are * stable enough today to floor (0.24% peak-to-peak at worst over five runs) and - * two of them — kotlin at 45.85 MiB and dart at 7.47 — are larger than budgeted + * two of them — kotlin at 40.82 MiB and dart at 7.47 — are larger than budgeted * arms, so a floor there would be worth having. That is a promotion to tier one, * with a ceiling and a recorded reading, and it is not this change: a floor * without them would assert "still measuring" against a number nothing else diff --git a/gitnexus/bench/kotlin-import-target/baselines.json b/gitnexus/bench/kotlin-import-target/baselines.json index 06bf6b633..d4c015be2 100644 --- a/gitnexus/bench/kotlin-import-target/baselines.json +++ b/gitnexus/bench/kotlin-import-target/baselines.json @@ -1,14 +1,15 @@ { - "_comment": "Baselines for bench/kotlin-import-target/measure.mjs --check. `fingerprint` is a sha256 over every `fileSet | fromFile | targetRaw -> result` record the correctness corpus resolves, in BOTH file-set iteration orders; it is a CORRECTNESS gate, so drift means Kotlin import resolution started returning a different file set and IMPORTS/CALLS edges moved in every Kotlin repository. Explain it, never re-baseline to make CI green. `cases` and `non_null` are asserted beside it because a shrunken or hollowed corpus produces a perfectly valid fingerprint over a smaller surface — all three are one re-baseline, never separate ones. `scaling_budget`, `depth_budget` and `small_ms_ceiling` are timing gates and carry deliberate headroom for shared CI runners.", - "_provenance": "RE-BASELINED ONCE, DELIBERATELY, IN #2881. The previous value ebf1790bf1d42dad483a51f2cbdeb2351e493b9e8236e4eedeef592dd81e2c5c (13256 non-null) was the PRE-INDEX implementation's, and the index that replaced it in #2872 reproduced it byte for byte — that is what made #2872 a performance change. #2881 changes resolution on purpose: `getKotlinFileIndex` no longer requires the parent directory to be the FIRST occurrence of that name in the path, so a file whose package directory name repeats higher in its own path is now a child of that package (`data/src/main/kotlin/com/example/data/Repo.kt` IS a child of `data`, and `import data.helper` resolves instead of returning null). The drift was not read off the new code and accepted; the corpus was dumped from both implementations and diffed record by record. Of 19968 distinct records, 235 moved: 54 NULL -> resolved (the fix itself, and exactly the +54 in non_null, 13256 -> 13310) and 181 answers that changed within a now-larger `dirChildren` bucket — 149 where the first-child tier picks a different member and 32 where a fan-out array gained entries. ZERO buckets lost a member, zero results were dropped, and in all 149 reselections the new answer's parent directory is the queried package segment. The corpus itself is untouched, which is why `cases` is unchanged at 20106 — the fingerprint is over the same surface as the value it replaces.", - "_gate_controls": "The gate is only worth its baseline if a plausible regression moves it, so each arm was checked against the mutation it exists to catch, with the resolver otherwise untouched. Caught against the CURRENT baseline (#2881): skipping the dirChildren suffix loop above depth 8 (fingerprint 41bb550b76d4…, non_null 13310 -> 12800); capping a dirChildren bucket at 17 entries (a7681945b752…, non_null unchanged — the fingerprint is the only arm that sees it); and, the mutation #2881 makes newly plausible, a HALF fix that drops only the `startsWith` guard and keeps the `indexOf` first-occurrence check (836977b83bf0…, non_null 13310 -> 13282), which leaves every mid-path repeat such as `top/data/mid/data/Repo.kt` still broken. Caught against the previous baseline and unaffected by the re-baseline: capping suffixByStem key depth at 7 (a0e6eb98f9…). Also caught, with the RESOLVER untouched and only the corpus edited: dropping the competing file from the exact-beats-earlier-suffix case and emptying the repeated-directory case (44df5093ee…). All of these passed silently before this corpus carried deep paths, packages above 16 files, queries against suffix keys deeper than 7, and the file set inside the hashed record. Re-check them after any corpus edit — a corpus that stops spanning an axis takes the gate with it.", + "_comment": "Baselines for bench/kotlin-import-target/measure.mjs --check. `fingerprint` is a sha256 over every `fileSet | fromFile | targetRaw -> result` record the correctness corpus resolves, in BOTH file-set iteration orders; it is a CORRECTNESS gate, so drift means Kotlin import resolution started returning a different file set and IMPORTS/CALLS edges moved in every Kotlin repository. Explain it, never re-baseline to make CI green. `cases` and `non_null` are asserted beside it because a shrunken or hollowed corpus produces a perfectly valid fingerprint over a smaller surface \u2014 all three are one re-baseline, never separate ones. `scaling_budget`, `depth_budget` and `small_ms_ceiling` are timing gates and carry deliberate headroom for shared CI runners.", + "_provenance": "RE-BASELINED ONCE, DELIBERATELY, IN #2881. The previous value ebf1790bf1d42dad483a51f2cbdeb2351e493b9e8236e4eedeef592dd81e2c5c (13256 non-null) was the PRE-INDEX implementation's, and the index that replaced it in #2872 reproduced it byte for byte \u2014 that is what made #2872 a performance change. #2881 changes resolution on purpose: `getKotlinFileIndex` no longer requires the parent directory to be the FIRST occurrence of that name in the path, so a file whose package directory name repeats higher in its own path is now a child of that package (`data/src/main/kotlin/com/example/data/Repo.kt` IS a child of `data`, and `import data.helper` resolves instead of returning null). The drift was not read off the new code and accepted; the corpus was dumped from both implementations and diffed record by record. Of 19968 distinct records, 235 moved: 54 NULL -> resolved (the fix itself, and exactly the +54 in non_null, 13256 -> 13310) and 181 answers that changed within a now-larger `dirChildren` bucket \u2014 149 where the first-child tier picks a different member and 32 where a fan-out array gained entries. ZERO buckets lost a member, zero results were dropped, and in all 149 reselections the new answer's parent directory is the queried package segment. The corpus itself is untouched, which is why `cases` is unchanged at 20106 \u2014 the fingerprint is over the same surface as the value it replaces.", + "_gate_controls": "The gate is only worth its baseline if a plausible regression moves it, so each arm was checked against the mutation it exists to catch, with the resolver otherwise untouched. All values below are against the CURRENT baseline (#2881, guards removed + per-directory key memo + bucket compaction). Caught: skipping the dirChildren component walk above depth 8 (fingerprint 41bb550b76d4\u2026, non_null 13310 -> 12800); capping a dirChildren bucket at 17 entries (a7681945b752\u2026, non_null UNCHANGED \u2014 the fingerprint is the only arm that sees it, and note the compaction pass now rewrites those same buckets, so this control was re-run after it); capping suffixByStem key depth at 7 (d24b8a2bd822\u2026, non_null unchanged); and a HALF fix that drops only the `startsWith` guard while keeping the `indexOf` first-occurrence check (836977b83bf0\u2026, non_null 13310 -> 13282), which leaves every mid-path repeat such as `top/data/mid/data/Repo.kt` broken and is the mutation #2881 itself makes plausible. Added with the memo: keying `dirKeys` on the directory's LAST SEGMENT instead of its full path (36a4e9dad313\u2026, non_null 13310 -> 13305) \u2014 the memo's whole safety argument is that its key determines the key SET a directory contributes, so a coarser key silently hands one directory another's bucket list, and that is the one way this optimization can move an answer. Also caught, with the RESOLVER untouched and only the corpus edited: dropping the competing file from the exact-beats-earlier-suffix case and emptying the repeated-directory case (44df5093ee\u2026). All of these passed silently before this corpus carried deep paths, packages above 16 files, queries against suffix keys deeper than 7, and the file set inside the hashed record. Re-check them after any corpus edit \u2014 a corpus that stops spanning an axis takes the gate with it. NOTE what no fingerprint control here can catch: the memo and the compaction are both invisible to this bench by design (identical output), so their correctness rests on a structural comparison of the three maps \u2014 key insertion order, bucket order, frozen-ness \u2014 over 1234 corpora in both iteration orders, not on this gate.", "fingerprint": "d91110bee389891c313811c5b4bae61d909561156e1458d38d487be969f0059c", "cases": 20106, "non_null": 13310, "scaling_budget": 1.6, - "depth_budget": 2.2, + "depth_budget": 2.0, "small_ms_ceiling": 40, - "_scaling_note": "(t_large/t_small)/(1600/400). ~1.0 is linear. OBSERVED BAND: 0.99-1.04 on a 12-core dev box, small arm ~6 ms. Read that band as a floor, not a spec — independent runs on other hardware during review came out 0.954-1.014, 0.965-1.036 and ~0.95-1.08, so a 1.2 reading is noise and should be re-run, not investigated. IMPORTS_PER_FILE is sized so the small arm lands in the ms rather than the ~2 ms a first revision measured, where timer granularity and JIT warm-up, not scaling, set the number; bench/cpp-qualified-ns documents the same artifact. TRIAGE: every timing arm here is a TIMING signal — RE-RUN IT on an idle machine before investigating; runner contention dominates. The fingerprint arm is the opposite: deterministic, a re-run never changes it, and it must never be wished away. FLOOR CHECK: the pre-index implementation — i.e. exactly the regression this gate exists to catch — measures ratio 3.737 on this corpus (2207.8 ms small, 33003.5 ms large, one cold run) against ~1.0 for the index. Independent review runs measured its floor at 3.905-4.297. Treat the absolute times as an order of magnitude only: the floor arm is one cold run because best-of-seven against a quadratic implementation costs minutes, while the index arm is best-of-seven after two warmups.", - "_depth_note": "deep_ms/shallow_ms at a FIXED file count, paths 24 components against 8. scaling_ratio divides the file count out, so it is scale-invariant and structurally cannot see a cost that grows with path depth instead — and both loops this change added are depth loops (one suffixByStem entry per '/' in a stem, one dirChildren pass per component of dir). OBSERVED BAND: 1.44-1.51 over four unloaded runs before #2881; 1.27-1.40 over five after. That change deleted two string comparisons per component of every dir, which is per-depth work, so the depth arm got cheaper while the file-count arm did not. The BUDGET moved with the band — 2.4 -> 2.2, holding the ~1.6x headroom over the band's top that 2.4 expressed against the old one; left at 2.4 it would have quietly become 1.8x, which is how a gate goes slack without anyone deciding to loosen it. Note what this budget is NOT for: a revert of #2881 scores ~1.5 and passes at either number, and that is correct — reverting it restores a resolution bug, which is the FINGERPRINT's job to catch, not a timing arm's. It sits above 1.0 legitimately: 3x the depth is 3x the suffix keys per file, so the build genuinely does more work; what the budget of 2.4 forbids is that growing faster than the depth ratio itself.", - "_ceiling_note": "small_ms_ceiling is an ABSOLUTE bound, because scaling_ratio is a ratio and a constant-factor regression that grows both arms equally passes it. Measured during review: a full workspace scan reintroduced on 1-in-16 imports is caught by the ratio (1.814), but at 1-in-32 it passes at 1.490 while running 2.8x slower in absolute terms. 40 ms against an observed 5.9-6.1 ms leaves ~6x of headroom for a loaded shared runner while still catching that shape." + "_scaling_note": "(t_large/t_small)/(1600/400). ~1.0 is linear. OBSERVED BAND: 0.99-1.04 on a 12-core dev box, small arm ~6 ms. Read that band as a floor, not a spec \u2014 independent runs on other hardware during review came out 0.954-1.014, 0.965-1.036 and ~0.95-1.08, so a 1.2 reading is noise and should be re-run, not investigated. IMPORTS_PER_FILE is sized so the small arm lands in the ms rather than the ~2 ms a first revision measured, where timer granularity and JIT warm-up, not scaling, set the number; bench/cpp-qualified-ns documents the same artifact. TRIAGE: every timing arm here is a TIMING signal \u2014 RE-RUN IT on an idle machine before investigating; runner contention dominates. The fingerprint arm is the opposite: deterministic, a re-run never changes it, and it must never be wished away. FLOOR CHECK: the pre-index implementation \u2014 i.e. exactly the regression this gate exists to catch \u2014 measures ratio 3.737 on this corpus (2207.8 ms small, 33003.5 ms large, one cold run) against ~1.0 for the index. Independent review runs measured its floor at 3.905-4.297. Treat the absolute times as an order of magnitude only: the floor arm is one cold run because best-of-seven against a quadratic implementation costs minutes, while the index arm is best-of-seven after two warmups.", + "_depth_note": "deep_ms/shallow_ms at a FIXED file count, paths 24 components against 8. scaling_ratio divides the file count out, so it is scale-invariant and structurally cannot see a cost that grows with path depth instead \u2014 and the two loops the index is built from are depth loops (one suffixByStem entry per '/' in a stem, one dirChildren pass per component of dir). OBSERVED BAND, five runs each on one box: 1.44-1.51 before #2881; 1.27-1.40 after its guard removal, which deleted two string comparisons per component of every dir; 1.20-1.26 after the same issue's per-directory key memo, which turns that whole component walk from once-per-FILE into once-per-DIRECTORY. Both movements are per-depth work, which is why this arm sees them and the file-count arm does not. The BUDGET moved with the band both times \u2014 2.4 -> 2.2 -> 2.0 \u2014 holding the ~1.6x headroom over the band's top that 2.4 expressed against the original; left at 2.4 it would quietly have become 1.9x, which is how a gate goes slack without anyone deciding to loosen it. Note what this budget is NOT for: a revert of #2881 scores ~1.5 and passes at any of those numbers, and that is correct \u2014 reverting it restores a resolution bug, which is the FINGERPRINT's job to catch, not a timing arm's. It sits above 1.0 legitimately: 3x the depth is 3x the suffix keys per file, so the build genuinely does more work; what the budget forbids is that growing faster than the depth ratio itself.", + "_ceiling_note": "small_ms_ceiling is an ABSOLUTE bound, because scaling_ratio is a ratio and a constant-factor regression that grows both arms equally passes it. Measured during review: a full workspace scan reintroduced on 1-in-16 imports is caught by the ratio (1.814), but at 1-in-32 it passes at 1.490 while running 2.8x slower in absolute terms. 40 ms against an observed 5.9-6.1 ms leaves ~6x of headroom for a loaded shared runner while still catching that shape.", + "_blind_spot": "WHAT THIS BENCH CANNOT SEE, measured rather than guessed. Its scaling corpus gives every module a UNIQUE package leaf (`com/example/mod{N}`), so a `dirChildren` query matches exactly one directory. That makes it blind to any cost that grows with the number of DIRECTORIES sharing a queried segment \u2014 the shape a real Kotlin monorepo has, where 200 modules each hold `data`, `ui` and `domain`. Established by building the reuse this file's memo argues against: swapping `dirChildren` for the shared `import-resolvers/package-dir-index.ts` (with its first-occurrence rule off) is OUTPUT-IDENTICAL \u2014 same fingerprint, same cases, same non_null, 0 divergences over 107948 answers \u2014 and on THIS corpus it costs only 1.37x-1.50x and passes every arm here. On a repeated-leaf corpus the same swap measures 13.5x per first-child query, 409x per fan-out, and 8114x on `import data.*` at 200 matching directories (it merges and SORTS every candidate, per import), for 3.1x-5.7x end to end and a bench-style scaling_ratio of 3.465 against this file's 1.6 budget \u2014 i.e. back to the pre-index quadratic floor of 3.737. A change that regresses this resolver to the very shape the bench exists to catch would go GREEN here. The trade it buys is real and also measured: 26.2% less retained memory, 12.18 MiB at 32000 files. If that memory is ever wanted, the shape to build is per-suffix keys -> DIRECTORY lists plus files-per-directory (8.29 MiB against 15.87 measured, single-directory query still one hash lookup) \u2014 and the corpus needs a repeated-leaf arm FIRST, or the gate will not be watching." } diff --git a/gitnexus/src/core/ingestion/languages/kotlin/import-target.ts b/gitnexus/src/core/ingestion/languages/kotlin/import-target.ts index 594525ba5..149ce57a7 100644 --- a/gitnexus/src/core/ingestion/languages/kotlin/import-target.ts +++ b/gitnexus/src/core/ingestion/languages/kotlin/import-target.ts @@ -188,6 +188,26 @@ const getKotlinFileIndex = perFileSet((allFilePaths: ReadonlySet): Kotli const exactByStem = new Map(); const suffixByStem = new Map(); const dirChildren: MutableDirChildren = new Map(); + /** + * BUILD-LOCAL: `dir` -> every `dirChildren` key a file in that directory + * contributes to. That list is a pure function of `dir`, and a package + * directory holds many files, so without this the walk below cuts one `slice` + * per component of the SAME directory once per FILE — and every slice after + * the first file's is a freshly allocated string that hashes to a key the map + * already holds and is then dropped. Interning them once per DIRECTORY + * instead of once per FILE is ~21% of the build at 32 000 files. + * + * It cannot move an answer. The array is filled on the first file of a + * directory, in the order the per-file walk produced, and every later file in + * that directory finds those keys already present — so the key set, the Map's + * key insertion order and every bucket's order are what the per-file form + * produced. Asserted structurally, not just through the fingerprint: the + * fingerprint only observes the index through the four resolver tiers, so a + * key-order move no corpus query reaches would survive it. + * + * Dropped with this frame, so it costs nothing retained. + */ + const dirKeys = new Map(); for (const raw of allFilePaths) { const norm = raw.replace(/\\/g, '/'); @@ -206,13 +226,18 @@ const getKotlinFileIndex = perFileSet((allFilePaths: ReadonlySet): Kotli if (!suffixByStem.has(suffix)) suffixByStem.set(suffix, raw); } - const lastSlash = norm.lastIndexOf('/'); - if (lastSlash < 0) continue; // repo-root file has no package directory - const dir = norm.slice(0, lastSlash); + // From `stem`, not `norm`: an extension carries no '/', so the last '/' of + // the two is the same character at the same index, and `stem.slice(0, + // dirEnd)` IS the string `norm.slice(0, norm.lastIndexOf('/'))` was. One + // backwards scan instead of two, over the string this loop already walked. + const dirEnd = stem.lastIndexOf('/'); + if (dirEnd < 0) continue; // repo-root file has no package directory + const dir = stem.slice(0, dirEnd); + const memo = dirKeys.get(dir); // The file's own directory always qualifies: the old scan's `atRoot` branch // matched `norm.startsWith(dir + '/')` and found no '/' after it. - addChild(dirChildren, dir, raw); + const keys = memo ?? [dir]; // Every component-suffix of the directory also qualifies: `s` is a suffix // of `dir` starting after a '/', so `dir` ends with `/s` by construction @@ -245,10 +270,14 @@ const getKotlinFileIndex = perFileSet((allFilePaths: ReadonlySet): Kotli // resolving to null on the very shape this fixes. The bucket is either // right or it is not; there is no version of it that is right for one // consumer and wrong for the other. - for (let i = 0; i < dir.length; i++) { - if (dir[i] !== '/') continue; - addChild(dirChildren, dir.slice(i + 1), raw); + if (memo === undefined) { + for (let i = 0; i < dirEnd; i++) { + if (dir[i] !== '/') continue; + keys.push(dir.slice(i + 1)); + } + dirKeys.set(dir, keys); } + for (const key of keys) addChild(dirChildren, key, raw); } // `findKotlinPackageFiles` hands a bucket straight out of the index — the @@ -260,7 +289,33 @@ const getKotlinFileIndex = perFileSet((allFilePaths: ReadonlySet): Kotli // reorder the cached bucket and flip the FIRST-child tier's answer for every // later import in the run. Freezing makes the contract true at runtime, so a // future mutation is a loud TypeError instead of a silent edge move. - for (const bucket of dirChildren.values()) Object.freeze(bucket); + // + // COMPACTED as they are frozen. `addChild` mints a bucket as `[raw]` and + // `push`es the rest, and V8 grows an array's backing store by + // `old + old/2 + 16`, so the SECOND child takes a 1-slot store to 17 and the + // eighteenth takes it to 41 — and every bucket then retains its overshoot for + // the life of the index. `slice()` allocates exactly `length`. Same fix and + // same accounting as the python `byBasename` sentence in + // `bench/import-target/baselines.json`: contents byte-identical, 5.13 MiB + // smaller at 32 000 paths, 11.2% of what that bench's kotlin heap arm reads + // (61 144 buckets x 88 B; 52.9% of those buckets' slots were empty). + // + // `length === 1` is skipped because a bucket that never grew is ALREADY exact + // — `[raw]` allocates one slot — so slicing it allocates a second array to + // save nothing. Not a micro-optimization: on a corpus of single-file packages + // the unguarded form costs 31% of the build and saves zero bytes. + // + // Re-`set` of a key the Map already holds keeps that key's position, so + // iteration order is unaffected. + for (const [key, bucket] of dirChildren) { + if (bucket.length === 1) { + Object.freeze(bucket); + continue; + } + const tight = bucket.slice(); + Object.freeze(tight); + dirChildren.set(key, tight); + } return { exactByStem, suffixByStem, dirChildren }; }); diff --git a/gitnexus/test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts b/gitnexus/test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts index ef38dd895..6da436625 100644 --- a/gitnexus/test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts +++ b/gitnexus/test/unit/scope-resolution/kotlin/kotlin-import-target-parity.test.ts @@ -135,7 +135,13 @@ describe('resolveKotlinImportTarget — index parity', () => { expect(resolve(['top/data/mid/data/Repo.kt'], 'data.something')).toEqual([ 'top/data/mid/data/Repo.kt', ]); - expect(resolve(['a/c/b/c/File.kt'], 'c.X')).toEqual(['a/c/b/c/File.kt']); + // `['a/c/b/c/File.kt'], 'c.X'` used to sit here too. It is the same shape + // with the segments renamed — four components, second and fourth equal, + // query the repeated name — so it could not fail while the case above + // passed. The bench corpus still carries it, where a second spelling of a + // shape costs nothing; a unit case that cannot distinguish two + // implementations is just a slower way to assert the first one. + // // Unrepeated control: the parent is the only occurrence. expect(resolve(['top/data/Repo.kt'], 'data.something')).toEqual(['top/data/Repo.kt']); });