This commit is contained in:
Gergő Magyar 2026-11-03 20:18:55 +00:00 • committed by GitHub
commit 7abc034f2a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
35 changed files with 49268 additions and 98 deletions

View file

@ -7,6 +7,84 @@ permissions:
contents: read
jobs:
lexical-lookup-benchmark:
name: lexical lookup / separate and repeated processes
runs-on: ubuntu-latest
timeout-minutes: 45
permissions:
contents: read
env:
GITNEXUS_LEXICAL_BENCH_WORKERS: '2'
GITNEXUS_LEXICAL_BENCH_HEAP_MIB: '4096'
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 2
persist-credentials: false
- name: Pin comparison revision
id: baseline
shell: bash
run: printf 'sha=%s\n' "$(git rev-parse HEAD^1)" >> "$GITHUB_OUTPUT"
- uses: ./.github/actions/setup-gitnexus
with:
build: 'true'
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
ref: ${{ steps.baseline.outputs.sha }}
path: baseline
persist-credentials: false
- name: Build baseline runtime and install search extensions
working-directory: baseline/gitnexus
run: |
npm ci
npm run build
npx tsx scripts/ensure-fts.ts
- name: Install candidate search extensions
working-directory: gitnexus
run: npx tsx scripts/ensure-fts.ts
- name: Prepare identical source corpora
shell: bash
env:
BASE_SHA: ${{ steps.baseline.outputs.sha }}
run: |
mkdir "$RUNNER_TEMP/lexical-lookup"
for variant in baseline candidate; do
corpus="$RUNNER_TEMP/lexical-lookup/corpus-$variant"
git clone --quiet --no-checkout --no-hardlinks "$GITHUB_WORKSPACE/baseline" "$corpus"
git -C "$corpus" sparse-checkout set gitnexus/src/core/ingestion gitnexus-shared/src/scope-resolution
git -C "$corpus" checkout --quiet --detach "$BASE_SHA"
done
- name: Compare lexical lookup semantics
run: node gitnexus/bench/lexical-lookup/differential.mjs "$GITHUB_WORKSPACE/baseline" "$GITHUB_WORKSPACE" | tee "$RUNNER_TEMP/lexical-lookup/differential.log"
shell: bash
- name: Measure separate processes
shell: bash
run: >-
node gitnexus/bench/lexical-lookup/measure.mjs
"$GITHUB_WORKSPACE/baseline" "$GITHUB_WORKSPACE"
"$RUNNER_TEMP/lexical-lookup/corpus-baseline" "$RUNNER_TEMP/lexical-lookup/corpus-candidate"
"$RUNNER_TEMP/lexical-lookup/fresh" fresh
2>&1 | tee "$RUNNER_TEMP/lexical-lookup/fresh.log"
- name: Measure repeated processes
if: ${{ !cancelled() }}
shell: bash
run: >-
node gitnexus/bench/lexical-lookup/measure.mjs
"$GITHUB_WORKSPACE/baseline" "$GITHUB_WORKSPACE"
"$RUNNER_TEMP/lexical-lookup/corpus-baseline" "$RUNNER_TEMP/lexical-lookup/corpus-candidate"
"$RUNNER_TEMP/lexical-lookup/persistent" persistent
2>&1 | tee "$RUNNER_TEMP/lexical-lookup/persistent.log"
- name: Retain lexical lookup measurements
if: always()
uses: ./.github/actions/upload-test-artifact
with:
name: lexical-lookup-performance
path: |
${{ runner.temp }}/lexical-lookup/fresh/results.json
${{ runner.temp }}/lexical-lookup/persistent/results.json
${{ runner.temp }}/lexical-lookup/*.log
retention-days: 30
mcp-query-benchmark:
name: MCP query performance
runs-on: ubuntu-latest

View file

@ -0,0 +1,91 @@
category,reader,repetition,count,extra_width_bytes,identity_bytes,force_gc,requested_calls,verified_calls,status,peak_rss_bytes,peak_rss_is_lower_bound,final_gc_heap_bytes,final_gc_rss_bytes
scale,original prepared async,1,0,0,0,true,3,3,complete,247959552,false,9354912,247320576
scale,cooperative prepared,1,0,0,0,true,3,3,complete,248381440,false,9348896,247791616
scale,cooperative prepared,1,16384,0,770048,true,3,3,complete,300908544,false,16922008,297074688
scale,original prepared async,1,16384,0,770048,true,3,3,complete,303788032,false,16924952,299225088
scale,original prepared async,1,32768,0,1540096,true,3,3,complete,322269184,false,24041632,318803968
scale,cooperative prepared,1,32768,0,1540096,true,3,3,complete,319725568,false,24046312,315506688
scale,cooperative prepared,1,65536,0,3080192,true,3,3,complete,354664448,false,37827224,349306880
scale,original prepared async,1,65536,0,3080192,true,3,3,complete,356151296,false,37828664,351764480
scale,original prepared async,1,131072,0,6160384,true,3,3,complete,418729984,false,64490336,410820608
scale,cooperative prepared,1,131072,0,6160384,true,3,3,complete,424943616,false,65540424,412405760
scale,cooperative prepared,1,262144,0,12320768,true,3,3,complete,558362624,false,121669088,528625664
scale,original prepared async,1,262144,0,12320768,true,3,3,complete,559734784,false,120741000,528830464
scale,original prepared async,1,524288,0,24641536,true,3,3,complete,818147328,false,232378984,764395520
scale,cooperative prepared,1,524288,0,24641536,true,3,3,complete,828125184,false,232392608,766697472
scale,cooperative prepared,1,1048576,0,49283072,true,3,3,complete,1357352960,false,455643288,1239048192
scale,original prepared async,1,1048576,0,49283072,true,3,3,complete,1362251776,false,454759344,1241427968
scale,cooperative prepared,2,1048576,0,49283072,true,3,3,complete,1356464128,false,454559560,1239228416
scale,original prepared async,2,1048576,0,49283072,true,3,3,complete,1373642752,false,455630704,1242320896
scale,original prepared async,2,524288,0,24641536,true,3,3,complete,830435328,false,232388752,770142208
scale,cooperative prepared,2,524288,0,24641536,true,3,3,complete,814604288,false,232378264,757653504
scale,cooperative prepared,2,262144,0,12320768,true,3,3,complete,555560960,false,121656056,528904192
scale,original prepared async,2,262144,0,12320768,true,3,3,complete,562339840,false,121638904,530608128
scale,original prepared async,2,131072,0,6160384,true,3,3,complete,418287616,false,64485216,410038272
scale,cooperative prepared,2,131072,0,6160384,true,3,3,complete,423247872,false,65547064,410828800
scale,cooperative prepared,2,65536,0,3080192,true,3,3,complete,355983360,false,37841384,351195136
scale,original prepared async,2,65536,0,3080192,true,3,3,complete,355446784,false,37824832,351186944
scale,original prepared async,2,32768,0,1540096,true,3,3,complete,322183168,false,24041472,317911040
scale,cooperative prepared,2,32768,0,1540096,true,3,3,complete,321650688,false,24042880,316248064
scale,cooperative prepared,2,16384,0,770048,true,3,3,complete,299732992,false,16924824,297156608
scale,original prepared async,2,16384,0,770048,true,3,3,complete,304021504,false,16916120,299577344
scale,original prepared async,2,0,0,0,true,3,3,complete,248582144,false,9354448,247795712
scale,cooperative prepared,2,0,0,0,true,3,3,complete,248983552,false,9348520,248360960
scale,original prepared async,3,0,0,0,true,3,3,complete,250396672,false,9354520,249204736
scale,cooperative prepared,3,0,0,0,true,3,3,complete,248913920,false,9348664,248324096
scale,cooperative prepared,3,262144,0,12320768,true,3,3,complete,557928448,false,121659672,530026496
scale,original prepared async,3,262144,0,12320768,true,3,3,complete,559079424,false,121618872,531521536
scale,original prepared async,3,32768,0,1540096,true,3,3,complete,320778240,false,24045264,316944384
scale,cooperative prepared,3,32768,0,1540096,true,3,3,complete,319979520,false,24038744,315535360
scale,cooperative prepared,3,1048576,0,49283072,true,3,3,complete,1356611584,false,455707776,1245691904
scale,original prepared async,3,1048576,0,49283072,true,3,3,complete,1358626816,false,454530440,1232388096
scale,original prepared async,3,65536,0,3080192,true,3,3,complete,356339712,false,37829552,351657984
scale,cooperative prepared,3,65536,0,3080192,true,3,3,complete,354988032,false,37827936,349990912
scale,cooperative prepared,3,524288,0,24641536,true,3,3,complete,824152064,false,232397976,769536000
scale,original prepared async,3,524288,0,24641536,true,3,3,complete,818388992,false,232381400,759271424
scale,original prepared async,3,16384,0,770048,true,3,3,complete,303210496,false,16928848,298729472
scale,cooperative prepared,3,16384,0,770048,true,3,3,complete,301977600,false,16920680,297541632
scale,cooperative prepared,3,131072,0,6160384,true,3,3,complete,426950656,false,65528616,412585984
scale,original prepared async,3,131072,0,6160384,true,3,3,complete,418197504,false,64493728,409317376
width,original prepared async,1,65536,128,28246016,true,3,3,complete,486682624,false,61827680,482648064
width,cooperative prepared,1,65536,128,28246016,true,3,3,complete,487243776,false,62853184,482398208
width,original prepared async,1,65536,512,103743488,true,3,3,complete,797528064,false,137084992,793079808
width,cooperative prepared,1,65536,512,103743488,true,3,3,complete,796639232,false,138111840,791740416
width,cooperative prepared,2,65536,128,28246016,true,3,3,complete,484167680,false,62862640,479428608
width,original prepared async,2,65536,128,28246016,true,3,3,complete,485789696,false,61833368,480915456
width,cooperative prepared,2,65536,512,103743488,true,3,3,complete,801923072,false,138116000,796282880
width,original prepared async,2,65536,512,103743488,true,3,3,complete,797220864,false,137084336,791736320
width,original prepared async,3,65536,128,28246016,true,3,3,complete,486744064,false,61823968,481890304
width,cooperative prepared,3,65536,128,28246016,true,3,3,complete,486674432,false,62855800,481607680
width,original prepared async,3,65536,512,103743488,true,3,3,complete,798748672,false,137085264,786731008
width,cooperative prepared,3,65536,512,103743488,true,3,3,complete,797835264,false,137993424,792526848
soak,original prepared async,,65536,0,3080192,true,40,40,complete,360906752,false,36883984,355561472
soak,cooperative prepared,,65536,0,3080192,true,40,40,complete,366809088,false,37905184,359317504
soak,cooperative prepared,,1048576,0,49283072,false,40,8,failed,2668290048,true,,
natural control,original prepared async,,1048576,0,49283072,false,12,12,complete,2331799552,false,453534800,1473449984
failure reproduction,cooperative prepared instrumented,,1048576,0,49283072,false,12,8,failed,2661085184,true,,
natural soak,cooperative direct,,1048576,0,49283072,false,40,40,complete,3041075200,false,454811184,2352312320
direct scale,cooperative direct,1,16384,0,770048,false,5,5,complete,316907520,false,16978936,312582144
direct scale,cooperative direct,1,65536,0,3080192,false,5,5,complete,430456832,false,37997512,406716416
direct scale,cooperative direct,1,131072,0,6160384,false,5,5,complete,583294976,false,65674696,522543104
direct scale,cooperative direct,1,262144,0,12320768,false,5,5,complete,846909440,false,121794512,725827584
direct scale,cooperative direct,1,524288,0,24641536,false,5,5,complete,1346248704,false,232843552,1101787136
direct scale,cooperative direct,1,1048576,0,49283072,false,5,5,complete,2418225152,false,454894384,1884545024
direct width,cooperative direct,1,65536,128,28246016,false,5,5,complete,642912256,false,62943728,601165824
direct width,cooperative direct,1,65536,512,103743488,false,5,5,complete,1121271808,false,138047664,1095659520
direct width,cooperative direct,2,65536,512,103743488,false,5,5,complete,1124130816,false,138171416,1099202560
direct width,cooperative direct,2,65536,128,28246016,false,5,5,complete,642789376,false,62951440,603402240
direct scale,cooperative direct,2,1048576,0,49283072,false,5,5,complete,2416967680,false,456022600,1898213376
direct scale,cooperative direct,2,524288,0,24641536,false,5,5,complete,1343852544,false,232557824,1100718080
direct scale,cooperative direct,2,262144,0,12320768,false,5,5,complete,845365248,false,121794224,724008960
direct scale,cooperative direct,2,131072,0,6160384,false,5,5,complete,583536640,false,65666832,521920512
direct scale,cooperative direct,2,65536,0,3080192,false,5,5,complete,430624768,false,38004072,407015424
direct scale,cooperative direct,2,16384,0,770048,false,5,5,complete,315084800,false,16971856,310923264
direct scale,cooperative direct,3,16384,0,770048,false,5,5,complete,313884672,false,16972968,309682176
direct scale,cooperative direct,3,65536,0,3080192,false,5,5,complete,430358528,false,37998152,406716416
direct scale,cooperative direct,3,131072,0,6160384,false,5,5,complete,583319552,false,65683272,522694656
direct scale,cooperative direct,3,262144,0,12320768,false,5,5,complete,844750848,false,121800720,723677184
direct scale,cooperative direct,3,524288,0,24641536,false,5,5,complete,1352495104,false,232659648,1105154048
direct scale,cooperative direct,3,1048576,0,49283072,false,5,5,complete,2419154944,false,456088936,1902014464
direct width,cooperative direct,3,65536,128,28246016,false,5,5,complete,644427776,false,62948408,604971008
direct width,cooperative direct,3,65536,512,103743488,false,5,5,complete,1108217856,false,138046296,1082867712
1 category reader repetition count extra_width_bytes identity_bytes force_gc requested_calls verified_calls status peak_rss_bytes peak_rss_is_lower_bound final_gc_heap_bytes final_gc_rss_bytes
2 scale original prepared async 1 0 0 0 true 3 3 complete 247959552 false 9354912 247320576
3 scale cooperative prepared 1 0 0 0 true 3 3 complete 248381440 false 9348896 247791616
4 scale cooperative prepared 1 16384 0 770048 true 3 3 complete 300908544 false 16922008 297074688
5 scale original prepared async 1 16384 0 770048 true 3 3 complete 303788032 false 16924952 299225088
6 scale original prepared async 1 32768 0 1540096 true 3 3 complete 322269184 false 24041632 318803968
7 scale cooperative prepared 1 32768 0 1540096 true 3 3 complete 319725568 false 24046312 315506688
8 scale cooperative prepared 1 65536 0 3080192 true 3 3 complete 354664448 false 37827224 349306880
9 scale original prepared async 1 65536 0 3080192 true 3 3 complete 356151296 false 37828664 351764480
10 scale original prepared async 1 131072 0 6160384 true 3 3 complete 418729984 false 64490336 410820608
11 scale cooperative prepared 1 131072 0 6160384 true 3 3 complete 424943616 false 65540424 412405760
12 scale cooperative prepared 1 262144 0 12320768 true 3 3 complete 558362624 false 121669088 528625664
13 scale original prepared async 1 262144 0 12320768 true 3 3 complete 559734784 false 120741000 528830464
14 scale original prepared async 1 524288 0 24641536 true 3 3 complete 818147328 false 232378984 764395520
15 scale cooperative prepared 1 524288 0 24641536 true 3 3 complete 828125184 false 232392608 766697472
16 scale cooperative prepared 1 1048576 0 49283072 true 3 3 complete 1357352960 false 455643288 1239048192
17 scale original prepared async 1 1048576 0 49283072 true 3 3 complete 1362251776 false 454759344 1241427968
18 scale cooperative prepared 2 1048576 0 49283072 true 3 3 complete 1356464128 false 454559560 1239228416
19 scale original prepared async 2 1048576 0 49283072 true 3 3 complete 1373642752 false 455630704 1242320896
20 scale original prepared async 2 524288 0 24641536 true 3 3 complete 830435328 false 232388752 770142208
21 scale cooperative prepared 2 524288 0 24641536 true 3 3 complete 814604288 false 232378264 757653504
22 scale cooperative prepared 2 262144 0 12320768 true 3 3 complete 555560960 false 121656056 528904192
23 scale original prepared async 2 262144 0 12320768 true 3 3 complete 562339840 false 121638904 530608128
24 scale original prepared async 2 131072 0 6160384 true 3 3 complete 418287616 false 64485216 410038272
25 scale cooperative prepared 2 131072 0 6160384 true 3 3 complete 423247872 false 65547064 410828800
26 scale cooperative prepared 2 65536 0 3080192 true 3 3 complete 355983360 false 37841384 351195136
27 scale original prepared async 2 65536 0 3080192 true 3 3 complete 355446784 false 37824832 351186944
28 scale original prepared async 2 32768 0 1540096 true 3 3 complete 322183168 false 24041472 317911040
29 scale cooperative prepared 2 32768 0 1540096 true 3 3 complete 321650688 false 24042880 316248064
30 scale cooperative prepared 2 16384 0 770048 true 3 3 complete 299732992 false 16924824 297156608
31 scale original prepared async 2 16384 0 770048 true 3 3 complete 304021504 false 16916120 299577344
32 scale original prepared async 2 0 0 0 true 3 3 complete 248582144 false 9354448 247795712
33 scale cooperative prepared 2 0 0 0 true 3 3 complete 248983552 false 9348520 248360960
34 scale original prepared async 3 0 0 0 true 3 3 complete 250396672 false 9354520 249204736
35 scale cooperative prepared 3 0 0 0 true 3 3 complete 248913920 false 9348664 248324096
36 scale cooperative prepared 3 262144 0 12320768 true 3 3 complete 557928448 false 121659672 530026496
37 scale original prepared async 3 262144 0 12320768 true 3 3 complete 559079424 false 121618872 531521536
38 scale original prepared async 3 32768 0 1540096 true 3 3 complete 320778240 false 24045264 316944384
39 scale cooperative prepared 3 32768 0 1540096 true 3 3 complete 319979520 false 24038744 315535360
40 scale cooperative prepared 3 1048576 0 49283072 true 3 3 complete 1356611584 false 455707776 1245691904
41 scale original prepared async 3 1048576 0 49283072 true 3 3 complete 1358626816 false 454530440 1232388096
42 scale original prepared async 3 65536 0 3080192 true 3 3 complete 356339712 false 37829552 351657984
43 scale cooperative prepared 3 65536 0 3080192 true 3 3 complete 354988032 false 37827936 349990912
44 scale cooperative prepared 3 524288 0 24641536 true 3 3 complete 824152064 false 232397976 769536000
45 scale original prepared async 3 524288 0 24641536 true 3 3 complete 818388992 false 232381400 759271424
46 scale original prepared async 3 16384 0 770048 true 3 3 complete 303210496 false 16928848 298729472
47 scale cooperative prepared 3 16384 0 770048 true 3 3 complete 301977600 false 16920680 297541632
48 scale cooperative prepared 3 131072 0 6160384 true 3 3 complete 426950656 false 65528616 412585984
49 scale original prepared async 3 131072 0 6160384 true 3 3 complete 418197504 false 64493728 409317376
50 width original prepared async 1 65536 128 28246016 true 3 3 complete 486682624 false 61827680 482648064
51 width cooperative prepared 1 65536 128 28246016 true 3 3 complete 487243776 false 62853184 482398208
52 width original prepared async 1 65536 512 103743488 true 3 3 complete 797528064 false 137084992 793079808
53 width cooperative prepared 1 65536 512 103743488 true 3 3 complete 796639232 false 138111840 791740416
54 width cooperative prepared 2 65536 128 28246016 true 3 3 complete 484167680 false 62862640 479428608
55 width original prepared async 2 65536 128 28246016 true 3 3 complete 485789696 false 61833368 480915456
56 width cooperative prepared 2 65536 512 103743488 true 3 3 complete 801923072 false 138116000 796282880
57 width original prepared async 2 65536 512 103743488 true 3 3 complete 797220864 false 137084336 791736320
58 width original prepared async 3 65536 128 28246016 true 3 3 complete 486744064 false 61823968 481890304
59 width cooperative prepared 3 65536 128 28246016 true 3 3 complete 486674432 false 62855800 481607680
60 width original prepared async 3 65536 512 103743488 true 3 3 complete 798748672 false 137085264 786731008
61 width cooperative prepared 3 65536 512 103743488 true 3 3 complete 797835264 false 137993424 792526848
62 soak original prepared async 65536 0 3080192 true 40 40 complete 360906752 false 36883984 355561472
63 soak cooperative prepared 65536 0 3080192 true 40 40 complete 366809088 false 37905184 359317504
64 soak cooperative prepared 1048576 0 49283072 false 40 8 failed 2668290048 true
65 natural control original prepared async 1048576 0 49283072 false 12 12 complete 2331799552 false 453534800 1473449984
66 failure reproduction cooperative prepared instrumented 1048576 0 49283072 false 12 8 failed 2661085184 true
67 natural soak cooperative direct 1048576 0 49283072 false 40 40 complete 3041075200 false 454811184 2352312320
68 direct scale cooperative direct 1 16384 0 770048 false 5 5 complete 316907520 false 16978936 312582144
69 direct scale cooperative direct 1 65536 0 3080192 false 5 5 complete 430456832 false 37997512 406716416
70 direct scale cooperative direct 1 131072 0 6160384 false 5 5 complete 583294976 false 65674696 522543104
71 direct scale cooperative direct 1 262144 0 12320768 false 5 5 complete 846909440 false 121794512 725827584
72 direct scale cooperative direct 1 524288 0 24641536 false 5 5 complete 1346248704 false 232843552 1101787136
73 direct scale cooperative direct 1 1048576 0 49283072 false 5 5 complete 2418225152 false 454894384 1884545024
74 direct width cooperative direct 1 65536 128 28246016 false 5 5 complete 642912256 false 62943728 601165824
75 direct width cooperative direct 1 65536 512 103743488 false 5 5 complete 1121271808 false 138047664 1095659520
76 direct width cooperative direct 2 65536 512 103743488 false 5 5 complete 1124130816 false 138171416 1099202560
77 direct width cooperative direct 2 65536 128 28246016 false 5 5 complete 642789376 false 62951440 603402240
78 direct scale cooperative direct 2 1048576 0 49283072 false 5 5 complete 2416967680 false 456022600 1898213376
79 direct scale cooperative direct 2 524288 0 24641536 false 5 5 complete 1343852544 false 232557824 1100718080
80 direct scale cooperative direct 2 262144 0 12320768 false 5 5 complete 845365248 false 121794224 724008960
81 direct scale cooperative direct 2 131072 0 6160384 false 5 5 complete 583536640 false 65666832 521920512
82 direct scale cooperative direct 2 65536 0 3080192 false 5 5 complete 430624768 false 38004072 407015424
83 direct scale cooperative direct 2 16384 0 770048 false 5 5 complete 315084800 false 16971856 310923264
84 direct scale cooperative direct 3 16384 0 770048 false 5 5 complete 313884672 false 16972968 309682176
85 direct scale cooperative direct 3 65536 0 3080192 false 5 5 complete 430358528 false 37998152 406716416
86 direct scale cooperative direct 3 131072 0 6160384 false 5 5 complete 583319552 false 65683272 522694656
87 direct scale cooperative direct 3 262144 0 12320768 false 5 5 complete 844750848 false 121800720 723677184
88 direct scale cooperative direct 3 524288 0 24641536 false 5 5 complete 1352495104 false 232659648 1105154048
89 direct scale cooperative direct 3 1048576 0 49283072 false 5 5 complete 2419154944 false 456088936 1902014464
90 direct width cooperative direct 3 65536 128 28246016 false 5 5 complete 644427776 false 62948408 604971008
91 direct width cooperative direct 3 65536 512 103743488 false 5 5 complete 1108217856 false 138046296 1082867712

File diff suppressed because one or more lines are too long

View file

@ -0,0 +1,156 @@
# Identity-reader memory scaling: evidence and limits
These measurements precede the production streaming implementation. See the
[follow-up fix and validation](identity-reader-fix-20261103.md) for its results.
The source-level live-memory bound is linear in graph nodes, stored identity
rows, and identity-string bytes. The measured process-memory results support
that growth rate over the tested range, but exposed a native allocation failure
in the first cooperative-reader probe. **That probe must not be retained.**
A separate direct-query routing probe passed the natural-GC soak and its own
size sweep. That array-based probe remained an isolated experiment: at the end
of this study no production adapter change had been made, and memory neutrality
had not been established.
## Main findings
| Check | Result |
| ---------------------------------------------------------- | ------------------------------------------------------------------------------------ |
| Cooperative reader using prepared execution, forced GC | Approximately linear RSS; R² 0.999875; held-out 1M result within 0.56% of prediction |
| Same reader, normal GC, 1,048,576 nodes | Native 512 MiB buffer pool exhausted on read 9; reproduced in a fresh process |
| Unchanged reader, same fixture/pool/heap, normal GC | All 12 requested reads passed |
| Cooperative reader using direct-query execution, normal GC | All 40 requested million-node reads passed |
| Direct-query probe, independent natural-GC size sweep | Approximately linear RSS; R² 0.998393; held-out 1M result 0.18% above prediction |
The failed reproduction recorded 249 cursor-close calls with no close errors.
Changing execution routing while retaining the SQL and row conversion avoided
the failure. This localizes the problem to the prepared-execution path in these
probes; it does not establish the exact native ownership or allocator bug.
## Normal-GC scaling of the direct-query probe
Each point below is the median of three fresh processes. Each process performs
five complete verifications, with no forced collection between them. The heap
limit is 4 GiB and the database buffer pool is fixed at 512 MiB. Collection runs
before measurement and once at the end to measure retained heap.
| Nodes | Peak RSS, MiB | Range, MiB |
| --------: | ------------: | --------------: |
| 16,384 | 300.5 | 299.3–302.2 |
| 65,536 | 410.5 | 410.4–410.7 |
| 131,072 | 556.3 | 556.3–556.5 |
| 262,144 | 806.2 | 805.6–807.7 |
| 524,288 | 1,283.9 | 1,281.6–1,289.8 |
| 1,048,576 | 2,306.2 | 2,305.0–2,307.1 |
Fitting only the points through 524,288 nodes gives:
`peak RSS ≈ 287.4 MiB + 2,014.7 bytes × node count`.
The held-out million-node prediction is 2,302.1 MiB; the observed median is
2,306.2 MiB, a 0.18% difference. Retained JavaScript heap also scales linearly:
about 424.7 bytes/node plus 9.7 MiB, R² 0.999995. The million-node retained heap
is 434.9 MiB, within 0.12% of that model's held-out prediction.
At 65,536 nodes, increasing total identity bytes from 2.94 to 26.94 to 98.94 MiB
increased median peak RSS from 410.5 to 613.1 to 1,069.3 MiB. The three-point
byte-size fit has R² 0.9961. It supports a byte-sensitive bound; it is not a
universal extrapolation to arbitrary string sizes or encodings.
## Repeated reads still have a cost
The direct-query probe's 40-read natural-GC run reached 2,900.2 MiB peak RSS.
Its first 12 reads reached 2,416.7 MiB, versus 2,223.8 MiB for the unchanged
reader's 12-read control: an observed **8.7% increase** in this single matched
comparison. This is not a replicated estimate of a general memory regression.
Median RSS rose from 2,427.0 MiB in reads 11–20 to 2,532.2 MiB in reads 31–40.
Final collection reduced live JavaScript heap to 433.7 MiB, but RSS remained
2,243.3 MiB. These results do not prove a permanent RSS plateau or leak freedom.
They establish that 40 reads completed without the prepared-path pool failure.
All 1,240 observed cursor closes succeeded; all 1,240 diagnostic prepared-object
finalizers had run by completion.
The smaller forced-GC 65K soaks were more stable: median retained RSS changed
by approximately 0.3 MiB for the original reader and 0.4 MiB for the cooperative
prepared reader between the same windows. That result did not predict the
larger normal-GC failure, which is why forced-GC evidence alone is insufficient.
## What the source proves
Let `N_e` be expected graph nodes, `N_d` stored rows scanned, and `S` their
combined identity-string size. These counts differ when the database contains
unexpected rows. The verifier can materialize those rows before rejecting them.
- Expected-node maps hold at most `N_e` entries, referencing existing nodes.
Normalized IDs add at most linear string storage.
- Each query returns one label's identity fields, with at most five scalar
fields per row. It does not project source content or relationships.
- The cooperative reader appends each row once to one array. It retains one
awaited continuation and no collection of previous arrays or promises.
- Validation keeps no additional graph-sized result structure. Maps and row
arrays belong to one invocation; the adapter closes cursors in `finally`.
Under the usual linear-storage assumption for V8 maps and dense arrays, the
bound on reachable JavaScript entries and payload is **O(N_e + N_d + S)**.
For a matching graph and fixed schema this is **O(N + S)**; fixed-width
identities reduce it to O(N). The existing input graph is additional and can
hold edges, source text, and other properties outside this claim.
The 1,024-row chunk controls yielding. It does **not** bound memory to 1,024
rows: the complete table result is retained. Reachable-object bounds do not
prove when V8 collects garbage, when native allocations are released, or how
much memory the OS reports as resident. The observed failure is a direct reason
not to equate the source-level bound with a total-process safety guarantee.
## Protocol and evidence
All reader probes used the unchanged PR runtime
`6aeeace8e96603fb062dcc709e834042d5a6df74`, Node 22.23.3, and LadybugDB 0.22.0.
Fixtures contain one Function table with unique fixed-width ASCII identities;
this exercises the largest possible table result for a given node count.
They contain no relationships or source content. Every completed verification
checks all identity fields against the in-memory graph, not just counts.
The initial controlled sweep covered zero nodes and seven doubling sizes from
16K through 1M, three processes per reader/size, and three reads per process.
Separate string-width cases and soak tests followed. The direct-query variant
then received its own prospectively recorded six-size/three-repeat natural-GC
protocol, wider-string cases, and five reads per process. Samples from the two
execution routes are kept separate.
The direct-query diagnostic still creates prepared objects for instrumentation
but executes the SQL through `Connection.query` instead of executing those
prepared objects. A production implementation has not been built or measured.
Prepared-object finalizer counts are observations, not proof of native ownership.
An initial 256 MiB-pool bulk load failed before any reader measurement. Setup
was then performed in separate 1 GiB-pool processes; every reported reader
measurement uses the same 512 MiB pool. Setup peaks are excluded. A planned
additional natural-GC sweep of the prepared path was cancelled after its soak
failed; its frozen plan is retained and is not counted as executed work.
The evidence comprises **90 reader processes: 88 completed and two failed on
read 9**, with **448 completed full verifications** across those processes.
These are repeated synthetic checks, not 448 distinct test cases. No native
worker was terminated. No application dependency or production source changed.
Matplotlib 3.11.2 was used in an isolated tool environment to render the figures.
- [Figure](identity-memory-scaling-20261103.png) · [Vector PDF](identity-memory-scaling-20261103.pdf)
- [Raw samples and exact probe sources](identity-memory-scaling-20261103.json)
- [Process-level CSV](identity-memory-scaling-20261103.csv)
The JSON includes source/driver hashes, protocols, all completed and failed
samples, failure messages, and original probe scripts. Each worker accepts a
JSON config containing `runtime`, `storage`, `output`, `count`, `width`, `mode`,
`variant`, `iterations`, and `forceGc`; it runs with
`node --expose-gc --max-old-space-size=4096 worker.mjs config.json`.
Use a disposable fixture directory and the pinned, built runtime. Run `seed`
in a separate process before `measure`; never target a production index.
The conclusion is limited to these fixtures and ranges. It is not a
full-indexing, watch-mode, arbitrary-input, or unlimited-repetition guarantee.
The original cooperative prepared path is rejected. These probes left the
direct-query route as a measured next candidate, with a memory tradeoff to
resolve before retention. The follow-up report above records that work.

Binary file not shown.

Binary file not shown.

After

Width:  |  Height:  |  Size: 349 KiB

View file

@ -0,0 +1,27 @@
case,reader,nodes,extra_bytes_per_string,reads,identity_bytes,peak_rss_mib,initial_heap_mib,retained_heap_mib,median_read_ms
soak-stream-1m-40,stream,1048576,0,40,49283072,2252.996094,433.200363,432.640045,2700.721579
baseline-1m-12,baseline,1048576,0,12,49283072,1972.113281,433.092522,432.438225,43542.687642
scale-r1-16384,stream,16384,0,5,770048,285.015625,15.861481,16.081467,50.761773
scale-r1-65536,stream,65536,0,5,3080192,335.550781,35.726349,35.942062,157.858161
scale-r1-131072,stream,131072,0,5,6160384,408.824219,62.143188,62.368248,310.738504
scale-r1-262144,stream,262144,0,5,12320768,545.914063,114.884155,115.117432,644.254398
scale-r1-524288,stream,524288,0,5,24641536,822.574219,221.255302,221.496635,1324.132947
scale-r1-1048576,stream,1048576,0,5,49283072,1375.195313,433.120430,433.388168,2745.030359
scale-r2-1048576,stream,1048576,0,5,49283072,1362.007813,434.165672,434.432190,2731.779330
scale-r2-524288,stream,524288,0,5,24641536,816.343750,221.280807,221.522964,1301.608581
scale-r2-262144,stream,262144,0,5,12320768,546.683594,114.899414,115.131882,613.970581
scale-r2-131072,stream,131072,0,5,6160384,408.570313,62.126694,62.352158,301.875663
scale-r2-65536,stream,65536,0,5,3080192,333.687500,35.727631,35.943748,151.271004
scale-r2-16384,stream,16384,0,5,770048,285.046875,15.860825,16.080383,49.948246
scale-r3-262144,stream,262144,0,5,12320768,547.566406,115.672760,115.906494,617.657477
scale-r3-16384,stream,16384,0,5,770048,286.113281,15.859398,16.079651,50.132029
scale-r3-1048576,stream,1048576,0,5,49283072,1354.042969,433.098320,433.365562,2688.166619
scale-r3-65536,stream,65536,0,5,3080192,336.796875,35.741119,35.960281,150.155641
scale-r3-524288,stream,524288,0,5,24641536,822.070313,221.264832,221.506134,1263.216974
scale-r3-131072,stream,131072,0,5,6160384,408.785156,62.141579,62.367043,295.693047
width-r1-128,stream,65536,128,5,28246016,445.675781,59.605537,59.822823,198.336899
width-r1-512,stream,65536,512,5,103743488,662.609375,131.409729,131.627014,321.027283
width-r2-128,stream,65536,128,5,28246016,445.824219,59.614334,59.831619,200.038170
width-r2-512,stream,65536,512,5,103743488,661.734375,131.443489,131.660637,323.354956
width-r3-128,stream,65536,128,5,28246016,443.546875,59.607437,59.825127,196.609874
width-r3-512,stream,65536,512,5,103743488,663.210938,131.439034,131.656319,322.631212
1 case reader nodes extra_bytes_per_string reads identity_bytes peak_rss_mib initial_heap_mib retained_heap_mib median_read_ms
2 soak-stream-1m-40 stream 1048576 0 40 49283072 2252.996094 433.200363 432.640045 2700.721579
3 baseline-1m-12 baseline 1048576 0 12 49283072 1972.113281 433.092522 432.438225 43542.687642
4 scale-r1-16384 stream 16384 0 5 770048 285.015625 15.861481 16.081467 50.761773
5 scale-r1-65536 stream 65536 0 5 3080192 335.550781 35.726349 35.942062 157.858161
6 scale-r1-131072 stream 131072 0 5 6160384 408.824219 62.143188 62.368248 310.738504
7 scale-r1-262144 stream 262144 0 5 12320768 545.914063 114.884155 115.117432 644.254398
8 scale-r1-524288 stream 524288 0 5 24641536 822.574219 221.255302 221.496635 1324.132947
9 scale-r1-1048576 stream 1048576 0 5 49283072 1375.195313 433.120430 433.388168 2745.030359
10 scale-r2-1048576 stream 1048576 0 5 49283072 1362.007813 434.165672 434.432190 2731.779330
11 scale-r2-524288 stream 524288 0 5 24641536 816.343750 221.280807 221.522964 1301.608581
12 scale-r2-262144 stream 262144 0 5 12320768 546.683594 114.899414 115.131882 613.970581
13 scale-r2-131072 stream 131072 0 5 6160384 408.570313 62.126694 62.352158 301.875663
14 scale-r2-65536 stream 65536 0 5 3080192 333.687500 35.727631 35.943748 151.271004
15 scale-r2-16384 stream 16384 0 5 770048 285.046875 15.860825 16.080383 49.948246
16 scale-r3-262144 stream 262144 0 5 12320768 547.566406 115.672760 115.906494 617.657477
17 scale-r3-16384 stream 16384 0 5 770048 286.113281 15.859398 16.079651 50.132029
18 scale-r3-1048576 stream 1048576 0 5 49283072 1354.042969 433.098320 433.365562 2688.166619
19 scale-r3-65536 stream 65536 0 5 3080192 336.796875 35.741119 35.960281 150.155641
20 scale-r3-524288 stream 524288 0 5 24641536 822.070313 221.264832 221.506134 1263.216974
21 scale-r3-131072 stream 131072 0 5 6160384 408.785156 62.141579 62.367043 295.693047
22 width-r1-128 stream 65536 128 5 28246016 445.675781 59.605537 59.822823 198.336899
23 width-r1-512 stream 65536 512 5 103743488 662.609375 131.409729 131.627014 321.027283
24 width-r2-128 stream 65536 128 5 28246016 445.824219 59.614334 59.831619 200.038170
25 width-r2-512 stream 65536 512 5 103743488 661.734375 131.443489 131.660637 323.354956
26 width-r3-128 stream 65536 128 5 28246016 443.546875 59.607437 59.825127 196.609874
27 width-r3-512 stream 65536 512 5 103743488 663.210938 131.439034 131.656319 322.631212

File diff suppressed because one or more lines are too long

View file

@ -0,0 +1,84 @@
# Identity reader memory fix — 2026-11-03
Status: implementation and validation complete. Local changes are not published.
PR #3805 remains at `6aeeace8e96603fb062dcc709e834042d5a6df74`. This follow-up starts at `9952270d81fc1f79bb8de34d43991b169d104289`, whose runtime is identical to that PR head. The compiled candidate hashes and complete measurements are in the accompanying JSON artifact. Node v22.23.3, LadybugDB 0.22.0, Linux x64.
## Change
Identity reconciliation consumes each row immediately through a direct native query. It no longer builds a JavaScript array for a complete table. The existing connection lock covers execution, synchronous row reads, cooperative yields every 1,024 rows, and cursor close. A read, validation or cleanup failure still prevents freshness publication. If validation and cleanup both fail, the original error is preserved.
Both checks remain in place: after COPY/checkpoint and immediately before freshness publication after the final checkpoint. The queries still scan each included label without filters and check every ID, name, path and source range, including unchanged nodes. Existing CSV normalization and exclusions are preserved. The general executeQuery/executePrepared/streamQuery behavior is unchanged.
Direct execution follows the evidence in [the preceding memory study](identity-memory-scaling-20261103.md): the prepared cooperative probe exhausted the fixed native pool on read 9 even with cursor closes; the direct-query probe completed 40 reads. This does not establish the precise C++ ownership defect or fix prepared execution globally.
## Memory evidence
All **172 complete identity verifications across 26 fresh reader processes passed**. Measurements use the same seeded fixtures, a 512 MiB native pool and a 4 GiB JavaScript heap limit. No GC is forced between reads. Initial/final GC is used only to measure retained memory. RSS includes native allocations and uses the OS high-water mark as well as periodic samples.
Two comparisons answer different questions:
- **Original PR implementation, matched first 12 million-node reads:** 1972.1 → 1926.7 MiB peak RSS (-2.3%). Median verification time: 43.54 → 2.68 seconds. This is one fresh paired memory comparison; an earlier original-reader run peaked at 2,223.8 MiB, so the small RSS difference should not be generalized beyond these runs.
- **Previous fast direct-query array probe, matched 40-read soak:** 2900.2 → 2253.0 MiB (-22.3%). The new reader completes all 40 million-node reads. Retained heap returns from 433.2 to 432.6 MiB after final GC. RSS rises and falls with normal GC; a flat RSS line is not claimed.
The following table compares the previous direct-query array probe with the implemented streaming reader. Each point is the median of three fresh processes, each performing five reads; all ranges are shown for the new reader.
| Nodes | Previous array probe peak MiB | Streaming peak MiB | Streaming range MiB | Change |
| --------: | ----------------------------: | -----------------: | ------------------: | -----: |
| 16,384 | 300.5 | 285.0 | 285.0–286.1 | -5.1% |
| 65,536 | 410.5 | 335.6 | 333.7–336.8 | -18.3% |
| 131,072 | 556.3 | 408.8 | 408.6–408.8 | -26.5% |
| 262,144 | 806.2 | 546.7 | 545.9–547.6 | -32.2% |
| 524,288 | 1283.9 | 822.1 | 816.3–822.6 | -36.0% |
| 1,048,576 | 2306.2 | 1362.0 | 1354.0–1375.2 | -40.9% |
Fit only through 524,288 nodes:
`peak RSS ≈ 268.12 MiB + 1,109.57 bytes × node count`
R² = 0.99994720. The held-out 1,048,576-node result is 1362.0 MiB versus 1377.7 MiB predicted (-1.14%). Retained heap fits approximately 9.38 MiB + 424.01 bytes/node, including the original graph.
Long identities are checked separately at 65,536 nodes. Adding 128 or 512 ASCII bytes to each ID/name/path gives 26.94 or 98.94 MiB of identity input and median RSS peaks of 445.7 and 662.6 MiB. Memory depends on identity bytes as well as node count.
![Memory scaling and repeated-read RSS](identity-reader-fix-20261103.png)
## What is established
The JavaScript live-data bound follows from the implementation: expected-node maps contain at most one entry per included graph node; each successful visit deletes one entry; the reader holds one decoded row and never accumulates rows. Auxiliary live JavaScript memory is O(N_expected + normalized expected-ID bytes + maximum row size). The original graph remains resident. There is no per-read global cache or retained result list.
Native query results still consume memory for the current label. The finite RSS measurements support approximately linear growth over 16,384–1,048,576 fixed-width nodes under the stated pool, heap and read count. They are not a mathematical proof about every native allocator, unlimited string widths, unlimited repetitions, or the full indexing pipeline. The 40-read soak separately checks repeated use; the held-out size and string-width cases guard against fitting only a convenient small range.
## End-to-end comparison
Three independent sessions each contain three paired edits in fresh processes and three in long-lived processes. Initial runtime order alternates between sessions, and edit order alternates within each session. Both runtimes use two parsing workers, a 4 GiB heap and FTS. The sparse 862-file corpus is pinned to `4916e0d5e1f4d6d70d41c1f39cd244bca2c7b390`. Each edit changes one file and uses incremental writes.
All 18 timed pairs have equal semantic graph digests and expected persisted call targets. Cold seeds are excluded from timing medians.
| Session | Process mode | Original median | Streaming median | Change |
| ------: | ------------ | --------------: | ---------------: | ------: |
| 1 | fresh | 19.164s | 17.356s | -9.43% |
| 1 | persistent | 18.397s | 16.569s | -9.94% |
| 2 | fresh | 19.144s | 17.377s | -9.23% |
| 2 | persistent | 18.504s | 16.805s | -9.18% |
| 3 | fresh | 20.554s | 17.635s | -14.20% |
| 3 | persistent | 20.385s | 17.766s | -12.85% |
Across all nine edit runs per mode, fresh-process medians are **19.164s → 17.435s (−9.02%)**, and long-lived-process medians are **18.826s → 16.805s (−10.73%)**. Median paired reductions are 9.23% and 10.76%, respectively. All 18 individual pairs improved, with visible variation; the host is shared. These results establish this corpus/workload, not a universal full-repository or watch-mode speedup.
Whole-process RSS high-water marks across timed samples are 1,947.6 → 1,937.7 MiB for fresh processes and 3,104.4 → 2,898.3 MiB for long-lived processes. The latter includes allocations from seeding and previous edits; these values must not be interpreted as isolated verifier memory.
## Validation
- Build, source typecheck and test typecheck passed. Changed-file ESLint completed with 0 errors and 44 warnings (including adapter/native-mock `any` types and an existing FTS-fixture non-null assertion).
- 39 focused tests passed: corruption rejection and recovery, CSV Unicode/null parity, row-at-a-time consumption, connection serialization across yields, cursor cleanup, incomplete/multiple results, and primary-error preservation.
- Native benchmark passed its normalized-scaling budget (0.713, budget 2), missing/swapped-field controls, and independent tuple oracle through selective DELETE/COPY, checkpoint and reopen.
- All 18 end-to-end timing pairs passed semantic graph and persisted-target checks.
- Full regression run: 31,574 passed, 103 skipped, five failures in one FTS crash-marker mock (4,788.40 s). The mock still wired the old row-array API; after adapting it to the visitor, all 38 tests in that file passed (50.04 s), as did test typechecking. The full run was not repeated after this test-only correction. Production sources and compiled runtime hashes remained identical to those measured.
- The local graph was refreshed. Final change analysis includes the new reader and completed without a partial or truncated result.
- Sequential source review found no actionable defect. No independent reviewer was dispatched under the user's AGENTS delegation mapping; this is not an independent review receipt.
## Reproduction
See [the benchmark instructions](../../gitnexus/bench/incremental-write-integrity/README.md#memory-scaling). The committed `memory.mjs` runs the actual compiled baseline or streaming implementation without prototype patches. Use separate processes and identical fixtures/limits for comparisons. The accompanying JSON preserves every raw run, fixture fingerprints, exact identity bytes, runtime hashes and per-read samples; CSV and PNG/PDF provide the plotted values.
[Raw evidence and scripts](identity-reader-fix-20261103.json) · [Process samples CSV](identity-reader-fix-20261103.csv) · [Figure PDF](identity-reader-fix-20261103.pdf)

Binary file not shown.

Binary file not shown.

After

Width:  |  Height:  |  Size: 221 KiB

View file

@ -0,0 +1,124 @@
# PR #3805 optimization assessment — 2026-11-03
No additional runtime optimization met the acceptance limits. The final runtime
is unchanged at `6aeeace8e96603fb062dcc709e834042d5a6df74`. The local
`optimize/lexical-3805` branch adds complete baseline collection, reversible
runtime ordering, regression tests, and measurement receipts.
## Measurements and decisions
Each session ran three real edits per lifecycle against matched unchanged-PR
controls. Cold seeds were excluded. A second session reversed both lifecycle
and initial runtime order. Values below are medians of session medians; with
two sessions, the outer median is their midpoint. Negative changes mean less
elapsed time.
| Candidate | Sessions | Fresh: reference → candidate | Change | Warm: reference → candidate | Change | Decision |
| --- | ---: | ---: | ---: | ---: | ---: | --- |
| Reuse claim buckets and specialize singleton selection | 1 | 29.799s → 32.233s | +8.17% | 34.280s → 33.238s | −3.04% | Rejected: fresh regression |
| Replace intermediate claim-selection arrays with scans | 2 | 26.347s → 25.941s | −1.54% | 23.526s → 23.491s | −0.15% | Inconclusive |
The limits were a repeatable warm-process improvement greater than 5%, at most
3% fresh-process regression, and three independent confirmation sessions before
any keep. Neither candidate reached confirmation eligibility. The second
candidate's first session improved neither objective; adding another session
cannot satisfy the configured requirement that every paired session improve
beyond the threshold. The decision helper returned terminal inconclusive after
the requested second sample. There is no confirmed additional speedup, and no
final before/after speedup is claimed for the unchanged runtime.
The [experiment receipt](../../gitnexus/bench/lexical-lookup/optimization-20261103.json)
contains all samples, graph hashes, call targets, memory/load diagnostics,
candidate patch hashes, forecasts, and decision outputs. It records 18 paired
edits across the two experiments: 36 measured analyses plus 12 cold seeds.
## Baseline identity and uncertainty
Both experiment runtimes used Node 22.23.3, LadybugDB 0.22.0, two workers and a
4 GiB heap. The sparse ingestion corpus was pinned to
`4916e0d5e1f4d6d70d41c1f39cd244bca2c7b390`. The host was shared; other workloads
were observable during measurement. The first large apparent warm gain in the
second experiment disappeared in its completed session median.
The earlier [PR-versus-base receipt](../../gitnexus/bench/lexical-lookup/baseline-20261103.json)
used PR `6b487a5` and base `4916e0d`, before the PR merged a Kotlin grammar
update. Its three independent sessions measured 28.494s → 27.552s fresh and
30.276s → 29.814s warm. Candidate session variation was 5.02% fresh and 17.01%
warm, so those small gains are inconclusive. Those older timings were never
used to score the newer experiments. The remote PR still pointed to `6aeeace8`
when checked during wrap-up.
Selection and reporting used the same sparse corpus. No holdout was configured,
and these results establish neither a full-repository nor a watch-mode speedup.
## What profiling justified trying
The unchanged PR's diagnostic CPU profile attributed about 7.8% of elapsed time
to lexical lookup, including claim selection. Garbage collection occupied 8.5%.
The CPU profile overlapped indexing, so it was used to locate costs, not to score
candidates. A separate allocation profile attributed approximately 1.6 GiB of
temporary allocations to claim selection, about 20% of sampled allocations.
These are cumulative allocation estimates, not live heap or peak RSS.
The first candidate's forecast was roughly 0.3–1.5s per 30s edit, plus unknown
GC savings. Its warm gain was smaller than the required threshold and fresh
time regressed. The second forecast was roughly 0.5–1.5s; its measured aggregate
warm reduction was only 0.035s. Allocation reduction alone did not establish an
elapsed-time benefit.
Remaining binding-filter and cycle-detection work has smaller attributed cost.
Its elapsed benefit remains unconfirmed and is unlikely to clear this run's 5%
threshold independently. A quieter benchmark host would help evaluate smaller
changes; broader pipeline changes require a separately scoped experiment.
## Correctness and retained changes
All 18 experiment edit pairs matched source/semantic graph hashes and persisted
call targets. Every edit changed the target between `save` and `load`; the runs
rejected no-op, skipped-FTS, collapsed-graph, or non-incremental shortcuts.
Derived communities/processes are excluded from the semantic digest, as in the
existing benchmark protocol.
The first candidate passed 3,472 scope tests and 72,000 differential comparisons.
The second passed 3,471 scope tests, 72,000 lookup comparisons, and 50,000
additional selector comparisons covering ordering, merges, hoisting, activation,
and non-finite positions. These counts describe separate runs and should not be
added as distinct tests. Both experimental source patches remain isolated.
The retained benchmark changes add `GITNEXUS_LEXICAL_BENCH_SCREEN=0` to retain
all pairs and `GITNEXUS_LEXICAL_BENCH_BASE_FIRST=0` to reverse initial ordering.
Defaults are preserved. Deterministic fresh/persistent regression tests verify
that even a slower candidate produces every requested sample and that temporary
control files are removed.
Final source and test typechecks, targeted tests, ESLint, formatting, and staged
diff checks passed. Complete graph change analysis reported 18 changed symbols
across six files, low risk, and no partial or truncated result. The benchmark's
external CI uses were also checked; an empty graph caller set was not treated
as proof that it was unused.
The full suite ran for 93 minutes: 31,567 tests passed, four failed, and 103
were skipped. All four failures were in unchanged Codex setup tests whose
temporary home directories were bypassed by the inherited `CODEX_HOME` setting.
Rerunning both affected files with child-process-only `env -u CODEX_HOME`
passed all 62 tests. The full suite was not repeated with that environment
correction; its original failure result remains in the validation receipt.
No final three-session timing confirmation was run: no runtime candidate
qualified for retention, and the remaining experiment budget could not fit
another confirmation cycle. Final elapsed-time values therefore remain
unconfirmed; unchanged runtime source is not a measured zero-percent gain.
## Run accounting
The optimization loop ran for 46 minutes 9 seconds. Kept: 0; rejected: 1;
inconclusive: 1; not selected: 0; censored: 0; errors/timeouts: 0. The remaining
hypothesis was deferred because it lacked a supported gain above the threshold
and the remaining approved time could not fit another full confirmation cycle.
No dependencies were added. No judges or subagents were used; monetary cost
telemetry was unavailable.
The resumable local ledger is
`.context/compound-engineering/ce-optimize/lexical-3805/experiment-log.yaml`,
with the approved `spec.yaml` beside it. The receipts linked above preserve the
measurements in the branch. No PR updates or pushes were made by this run.

View file

@ -0,0 +1,120 @@
# PR #3805: identity-read optimization experiment
This report records the rejected batching experiment. The subsequent streaming
implementation and its validation are tracked in the
[identity-reader fix report](identity-reader-fix-20261103.md).
The additional batching candidate was rejected because it doubled peak resident
memory in the controlled large-table probe. No runtime change was retained.
At that decision, the runtime remained `6aeeace8e96603fb062dcc709e834042d5a6df74`; this report and
the [measurement receipt](pipeline-3805-results.json) preserve the evidence.
## Matched baseline and candidate
Two independent sessions each measured three real edits in fresh processes and
three in a persistent process. Session two reversed lifecycle and initial runtime
order. Cold seeds were excluded. Every measured pair matched the semantic graph
and persisted call targets; all edits used incremental writes.
| Session | Fresh reference → batch | Warm reference → batch |
| ------------------------- | ----------------------- | ---------------------- |
| 1 | 22.182s → 21.639s | 22.827s → 20.864s |
| 2 | 23.181s → 21.174s | 19.852s → 18.743s |
| Median of session medians | 22.682s → 21.406s | 21.339s → 19.803s |
The aggregate reductions were 1.275s (5.62%) fresh and 1.536s (7.20%) warm.
The forecast was approximately 1.5–2.5s per 24-second warm edit; the measured
warm reduction is near its lower bound, but cannot qualify as a retained gain.
The required third session was not run after the memory failure. These are
diagnostic candidate results, not confirmed improvements to the final branch.
Both runtimes used Node 22.23.3, LadybugDB 0.22.0, two parsing workers, and a
4 GiB heap. The sparse ingestion corpus was pinned to
`4916e0d5e1f4d6d70d41c1f39cd244bca2c7b390`. The comparison reference was the
unchanged PR runtime; candidate patch SHA-256 was
`8350e17daba98dfbf3042ab233ae7b9c2b924cfcfaa2ccdd979bcb9795812a25`.
The host is shared, and session variation is visible in the table. No holdout
was configured. No full-repository or watch-mode improvement is established.
## Why batching was rejected
The candidate kept both full identity checks and collected complete identity
structs per table, avoiding one asynchronous native call per row. All 33 targeted
tests passed, including corruption detection at both publication boundaries.
Build, source/test typechecks, and changed-file ESLint passed. The native
16K/65K-node benchmark also passed missing/swapped-field controls and checked
selective delete/COPY, checkpoint, and reopen against an independent tuple oracle.
Timing and graph equality did not expose the memory cost. A separate probe used
65,536 Function nodes, a fixed 256 MiB database buffer pool, explicit GC, four
verifier calls, and separate reference/candidate processes on the same seeded
database. Reference peak RSS was 346,552 KiB (338.4 MiB); unbounded collection
reached 709,272 KiB (692.6 MiB). Retained JavaScript heap remained about 30 MB in
both. The native allocation mechanism was not established.
Two scratch alternatives were also unsuitable. Ordered paging failed closed on
a duplicate identity in the healthy fixture; no engine root cause is claimed.
Grouping identities by physical offset preserved the fixture but still reached
675,416 KiB. Neither was incorporated into runtime source.
The correctness review found the measured RSS regression. It was sequential;
the independent adversarial lane did not run under the repository's solo mapping.
Its verdict is **Not ready**. The full suite was not rerun for this rejected
candidate. The earlier unchanged-runtime full-suite result and its setup-test
environment limitation remain documented in [the prior report](lexical-3805-report.md).
## Next measured opportunity
The installed driver exposes `getNextSync()`. A process-local probe retained the
original unfiltered queries and consumed rows synchronously, yielding with
`setImmediate` every 1,024 rows. No installed dependency or runtime source was
modified by this probe.
| Controlled 65K-node probe | Original reader | Cooperative synchronous reader |
| ----------------------------- | --------------: | -----------------------------: |
| Median of four verifier calls | 3,395.766ms | 166.648ms |
| Peak RSS | 346,552 KiB | 348,804 KiB |
The standalone native benchmark measured 50.930ms at 16K nodes and 146.986ms at
65K nodes. Both sizes passed the independent oracle after selective writes,
checkpointing, and read-only reopening, and rejected missing/swapped fields.
These are local verifier measurements, not end-to-end timing confirmation.
A next implementation should make this reader opt-in at the two identity-check
call sites, preserving existing connection locking, error propagation, cursor
closing, and default query behavior. It needs yielding/error/cleanup tests,
larger-table memory and responsiveness checks, then three matched fresh/warm
sessions and the full suite. No new dependency is needed.
Subsequent [memory-scaling tests](identity-memory-scaling-20261103.md) found that
the cooperative prepared path exhausts the native pool on its ninth repeated
million-node read without forced GC. A separate direct-query probe passed 40
reads and its natural-GC size sweep, but showed higher peak RSS than the original
reader. The prepared-path proposal above is therefore superseded by that evidence.
Fresh graph impact for the shared `readQueryRows` helper is **CRITICAL**:
27 affected symbols, six direct callers and six processes, including query
handling, embedding reads, statistics, and analysis. This expands the approved
mutable scope beyond the reconciliation module and requires maintainer sign-off
under GUARDRAILS.md. A global replacement is not proposed.
## Run state
The one approved implementation experiment completed and was rejected; kept 0,
rejected 1, not selected 0, inconclusive 0, censored 0, experiment errors/timeouts 0.
Scratch paging failure is recorded separately above. The optimization loop took
about 30 minutes before wrap-up; the durable ledger records exact active time.
No judges, subagents, new dependencies, pushes, or PR comments were used. Monetary
cost telemetry was unavailable.
At this experiment's closure, no final speedup or measured zero-percent result
was claimed for the unchanged runtime. The greater-than-5% warm target remained
unmet by a retained change; fresh timing likewise had no new confirmed value.
The adapter opportunity was measured but unimplemented. Its subsequent work is
recorded separately in the follow-up report linked above.
Resumable state is under
`.context/compound-engineering/ce-optimize/pipeline-3805/`, including
`experiment-log.yaml`, `spec.yaml`, archived candidate patch and raw measurements.
The tracked receipt beside this report preserves the paired timings, graph
checks, memory samples, native results, and impact report outside that local state.

File diff suppressed because it is too large Load diff

View file

@ -29,11 +29,15 @@ export interface NameClaimSources {
readonly bindingsAt?: (scope: Scope, name: string) => readonly BindingRef[];
readonly importsAt?: (scope: Scope) => readonly ImportEdge[];
}
const NO_BINDINGS: readonly BindingRef[] = Object.freeze([]);
const NO_CLAIMS: readonly NameClaim[] = Object.freeze([]);
const NO_IMPORTS: readonly ImportEdge[] = Object.freeze([]);
const NO_SELECTION: ScopeClaimSelection = Object.freeze({ status: 'absent', claims: NO_CLAIMS });
const ABSENT: NameClaimResult = Object.freeze({
status: 'absent',
bindings: [],
imports: [],
claims: [],
bindings: NO_BINDINGS,
imports: NO_IMPORTS,
claims: NO_CLAIMS,
});
// Published scope arrays are readonly. Cache by the array, not the scope ID:
// finalization and cache hydration can replace a scope's import/claim snapshot.
@ -42,8 +46,6 @@ const importsByName = new WeakMap<
readonly ImportEdge[],
ReadonlyMap<string, readonly ImportEdge[]>
>();
const NO_CLAIMS: readonly NameClaim[] = Object.freeze([]);
const NO_IMPORTS: readonly ImportEdge[] = Object.freeze([]);
export function nameClaimsFor(scope: Scope, name: string): readonly NameClaim[] {
const claims = scope.nameClaims;
@ -93,11 +95,13 @@ export function selectClaimsAtScope(
name: string,
options: NameLookupOptions = {},
): ScopeClaimSelection {
const candidates = nameClaimsFor(scope, name);
if (candidates.length === 0) return NO_SELECTION;
const purpose = options.purpose ?? 'value';
const claims = nameClaimsFor(scope, name).filter(
const claims = candidates.filter(
(c) => c.purpose === undefined || c.purpose === 'both' || c.purpose === purpose,
);
if (claims.length === 0) return { status: 'absent', claims };
if (claims.length === 0) return NO_SELECTION;
const active = claims.filter(
(c) =>
c.availableFrom === undefined ||
@ -192,6 +196,9 @@ export function lookupLexicalName(
): NameClaimResult {
let id: ScopeId | null = startScope;
let position = options.position;
// Position is cleared when crossing deferred/redirected ownership. Reuse
// the caller's readonly options until that happens, then copy only once.
let selectionOptions = options;
let forcedClaims: readonly NameClaim[] | undefined;
let carriedConstructor: TypeRef | undefined;
const visited = new Set<ScopeId>();
@ -209,9 +216,10 @@ export function lookupLexicalName(
id = scope.lookupPolicy?.parentScope ?? scope.parent;
continue;
}
if (selectionOptions.position !== position) selectionOptions = { ...options, position };
const selected: ScopeClaimSelection =
forcedClaims === undefined
? selectClaimsAtScope(scope, name, { ...options, position })
? selectClaimsAtScope(scope, name, selectionOptions)
: { status: 'selected', claims: forcedClaims };
forcedClaims = undefined;
const redirected = selected.claims.find((c) => c.redirect !== undefined);
@ -245,38 +253,47 @@ export function lookupLexicalName(
return { ...ABSENT, status: 'blocked', scope, claims: selected.claims };
const nameClaims = nameClaimsFor(scope, name);
const hasClaimsForName = nameClaims.length > 0;
const allBindings = sources.bindingsAt?.(scope, name) ?? scope.bindings.get(name) ?? [];
const allBindings =
sources.bindingsAt?.(scope, name) ?? scope.bindings.get(name) ?? NO_BINDINGS;
const allImports = importsFor(sources.importsAt?.(scope) ?? scope.imports ?? NO_IMPORTS, name);
let bindings = allBindings.filter(
(b) =>
b.availableFrom === undefined ||
position === undefined ||
compare(b.availableFrom, position) <= 0,
);
let bindings =
allBindings.length === 0
? NO_BINDINGS
: allBindings.filter(
(b) =>
b.availableFrom === undefined ||
position === undefined ||
compare(b.availableFrom, position) <= 0,
);
let imports = allImports;
if (selected.status === 'selected') {
bindings = bindings.filter((b) =>
selected.claims.some((c) =>
c.kind === 'import'
? b.via !== undefined && matchesClaim(b.via.atRange, c)
: c.kind === 'binding' &&
(b.origin === 'local' || b.declarationRange !== undefined) &&
matchesBindingClaim(b.declarationRange, c),
),
);
imports = allImports.filter((e) =>
selected.claims.some((c) => c.kind === 'import' && matchesClaim(e.atRange, c)),
);
if (bindings.length > 0)
bindings = bindings.filter((b) =>
selected.claims.some((c) =>
c.kind === 'import'
? b.via !== undefined && matchesClaim(b.via.atRange, c)
: c.kind === 'binding' &&
(b.origin === 'local' || b.declarationRange !== undefined) &&
matchesBindingClaim(b.declarationRange, c),
),
);
imports =
allImports.length === 0
? NO_IMPORTS
: allImports.filter((e) =>
selected.claims.some((c) => c.kind === 'import' && matchesClaim(e.atRange, c)),
);
} else if (hasClaimsForName) {
// A declaration in another namespace, or a not-yet-visible declaration,
// must not sneak back through the unfiltered legacy binding channels.
bindings = bindings.filter(
(b) =>
b.availableFrom !== undefined &&
b.declarationRange !== undefined &&
!nameClaims.some((c) => rangesOverlap(c.range, b.declarationRange!)),
);
imports = [];
if (bindings.length > 0)
bindings = bindings.filter(
(b) =>
b.availableFrom !== undefined &&
b.declarationRange !== undefined &&
!nameClaims.some((c) => rangesOverlap(c.range, b.declarationRange!)),
);
imports = NO_IMPORTS;
}
const ownType = scope.typeBindings.get(name);
const typeFactOwnsName =
@ -328,9 +345,12 @@ export function lookupLexicalName(
(typeBinding === undefined || isInferredFact(typeBinding.source))
)
typeBinding = carriedConstructor;
const usable = bindings.filter(
(b) => (options.purpose ?? 'value') !== 'value' || b.via?.typeOnly !== true,
);
const usable =
bindings.length === 0
? NO_BINDINGS
: bindings.filter(
(b) => (options.purpose ?? 'value') !== 'value' || b.via?.typeOnly !== true,
);
return {
status:
usable.length > 0 ||
@ -358,6 +378,8 @@ export function lookupLexicalName(
}
// A write to an undeclared name declares nothing, but it still carries the
// value's type to reads inside the writing scope.
const carried = carriedConstructor === undefined ? {} : { typeBinding: carriedConstructor };
if (carriedConstructor === undefined)
return id === null ? ABSENT : { ...ABSENT, status: 'blocked' };
const carried = { typeBinding: carriedConstructor };
return id === null ? { ...ABSENT, ...carried } : { ...ABSENT, ...carried, status: 'blocked' };
}

View file

@ -1,7 +1,6 @@
# Analyze phase breakdown — where the time actually goes
Updated 2026-11-01 with scope-emission measurements after #3796 merged, plus
production-hardening measurements on LadybugDB **0.22.0** for unchanged dirty checkouts.
Updated 2026-11-02 with the lexical-lookup allocation experiment below.
Records what the `analyze`
pipeline costs per phase and the results of optimization experiments.
The September measurements remain below as historical
@ -14,6 +13,192 @@ dispatch path is `bench/parse-dispatch-rounds/`.
---
## Lexical-lookup allocation experiment — 2026-11-02
**Decision: keep experimental; the performance requirement for both workflows
has not passed.** Fresh-process edits improved in all three pairs (8.95% lower
median), but the first repeated-process pair took 12.68% longer and triggered
the harness's early rejection gate. Its host load rose sharply. That does not
prove a patch-induced regression, and it also does not establish a warm-process
speedup. Do not promote this as a production optimization on this evidence.
The patch reduces temporary allocation in `lookupNameClaim`,
`selectClaimsAtScope` and `lookupLexicalName`: reuse frozen empty results, skip
filters over empty inputs, avoid concatenating empty binding channels, and copy
lookup options only when the effective position changes. Nonempty binding
results retain snapshot semantics. Provider hooks, declaration ownership,
shadowing, redirects and position checks still run. No semantic-result cache,
new invalidation mechanism or native dependency is introduced.
### Separate-process edit measurements
| Measurement | Baseline | Candidate | Less elapsed time |
| --- | ---: | ---: | ---: |
| Edit 1: `save` → `load` | 122.065s | 111.135s | 8.95% |
| Edit 2: `load` → `save` | 114.006s | 101.730s | 10.77% |
| Edit 3: `save` → `load` | 131.285s | 111.303s | 15.22% |
| Median edit | 122.065s | 111.135s | 8.95% |
| All three edits | 367.356s | 324.168s | 11.76% |
| Median scope-resolution phase | 68.197s | 62.344s | 8.58% |
| Highest process peak RSS across edits | 4.29 GiB | 4.25 GiB | — |
All three scope-resolution phases improve, but not all wall-time differences
belong to that phase. For example, the first pair's process-detection interval
falls from 5.889s to 1.475s, and the third pair's post-`done` work falls from
16.551s to 7.974s. The patch does not directly change either stage. Treat the
observed end-to-end reductions as a small shared-host screen, not a universal
speedup or a precise attribution of saved time. One-minute host load ranges
from 7.79 to 11.43 at the start of these measured edits.
### Repeated-process screen
| Measurement | Baseline | Candidate | Less elapsed time |
| --- | ---: | ---: | ---: |
| First edit: `save` → `load` | 109.483s | 123.364s | −12.68% |
| Scope-resolution phase | 67.013s | 65.739s | 1.90% |
| Ladybug write intervals | 8.814s | 12.902s | — |
| FTS phase | 10.600s | 14.766s | — |
| Work after `done` progress | 8.947s | 14.858s | — |
| Process peak RSS, including setup | 5.49 GiB | 5.62 GiB | — |
These intervals come from progress callbacks; they include all work between
adjacent callbacks rather than instrumenting native operations individually.
Baseline one-minute load fell **6.35 → 3.81**; candidate load rose
**3.81 → 12.23**. Scope resolution improved slightly while several other
stages slowed. Only this first pair was run, as specified by the screening
gate; there is no three-pair warm-process median. Both graphs and persisted
call targets matched. Persistent mode retains one process per variant and
models repeated analysis, not watch-event scheduling or debounce latency.
Separate excluded cold setup runs took 218.219s/187.148s in fresh mode and
158.148s/153.951s in persistent mode (baseline/candidate). Each lifecycle uses
independent index directories. Setup order is fixed and host load changes,
so these are not controlled cold-index speedup claims. The
[measurement receipt](lexical-lookup/measurements.json) includes every sample,
phase duration, peak memory, graph hash, load average and runtime checksum.
### Method and correctness boundaries
Baseline runtime and both disposable corpus checkouts are pinned to
`347cc99b4a237f9a71c0084167bcbcbd9d06e610`; the candidate runtime adds only the
lookup patch. Each corpus has 3,384 indexed files / 3,120 parseable files,
including the same control file. Node **v22.23.3**, LadybugDB **0.22.0**,
Linux x64, Intel i9-12900K (24 logical CPUs), 16 parse workers, 8 GiB V8 heap.
Samples run serially; paired order is baseline/candidate, candidate/baseline,
baseline/candidate after excluded initialization runs. Runtime builds stay
fixed throughout each comparison, with no tests or other benchmarks launched
by this experiment concurrently. The machine is shared with other workloads.
The [harness](lexical-lookup/measure.mjs) times `runFullAnalysis`, including
incremental persistence and FTS. Process startup, graph hashing and subsequent
persisted-target verification are outside the timer. CPU profiling and heap
probes are off; the existing scope phase logger remains on for both variants.
Each measured edit changes one file, reparses 22 files, hits the parser cache
for 3,098 files, uses incremental writes and has no rebuild reasons. FTS must
remain available. This is a real edit benchmark, not an up-to-date fast return.
Every pair compares canonical hashes of **114,413 source/semantic nodes and
318,798 relationships**. Derived Folder/Community/Process nodes and
CONTAINS/MEMBER_OF/STEP_IN_PROCESS edges are excluded. Hashes cover the graph
returned after persistence; cleared source content is not independently
verified by these hashes. A separate LadybugDB query after each sample verifies
that the control CALLS edge follows `save`/`load`. Matching graphs preserve the
baseline's limits; they do not establish complete resolution of the corpus.
Targeted tests cover mutation after a missing lookup, result snapshots, reused
frozen options and deferred activation. The initial validation passed 137 tests
across lexical ownership, TypeScript type ownership, block shadowing and
TypeScript/Python/C++ native rename suites. A seeded differential screen also
matched the original implementation in 72,000 generated lexical lookups.
The broader scope-resolution suite then passed **3,467 tests / 175 files**
(overlaps the initial targeted tests). The parameterized differential harness
also passed when rerun against the same two builds.
Build, test typecheck and ESLint passed (existing non-null assertion warnings
remain). These were the checks completed alongside the benchmark; PR validation is recorded separately.
To reproduce, build the baseline and candidate runtimes first, prepare two
clean disposable corpus checkouts at the same commit, then run from
`gitnexus/` (the runtime arguments name repository roots):
```sh
node bench/lexical-lookup/differential.mjs BASE_RUNTIME CANDIDATE_RUNTIME
node bench/lexical-lookup/measure.mjs BASE_RUNTIME CANDIDATE_RUNTIME BASE_CORPUS CANDIDATE_CORPUS OUTPUT_FRESH fresh
node bench/lexical-lookup/measure.mjs BASE_RUNTIME CANDIDATE_RUNTIME BASE_CORPUS CANDIDATE_CORPUS OUTPUT_PERSISTENT persistent
```
By default, the timing harness stops after the first edit pair if the candidate
is not faster. To establish a baseline, set `GITNEXUS_LEXICAL_BENCH_SCREEN=0`
to collect all three pairs regardless of their timings. Repeat each lifecycle
in independent sessions, setting `GITNEXUS_LEXICAL_BENCH_BASE_FIRST=0` in
alternating sessions to reverse seed and edit ordering. Both settings accept
only `0` or `1` and are recorded in `results.json`. Report every sample and
the variation between sessions; a screening result alone is not a baseline.
The harness waits for native workers to exit normally and removes only its
control source file. Keep builds fixed and do not run tests concurrently.
The receipt's harness checksums identify the versions used for measurement.
The checked-in scripts were subsequently formatted, and cleanup now attempts
both worker shutdowns and removes the control files even if a worker fails.
These changes are outside the timed analysis interval.
### CI execution
`.github/workflows/ci-tests.yml` runs the differential check and both process
lifecycles in a dedicated job. It builds the candidate and its first parent
(the base branch on a pull-request merge checkout), then uses two clean sparse
corpus checkouts at that parent. The CI corpus contains the ingestion and shared
scope-resolution source trees, plus Git's cone-mode parent files. It is smaller
than the full-repository corpus measured above; its timings are a separate series.
CI sets `GITNEXUS_LEXICAL_BENCH_WORKERS=2` and
`GITNEXUS_LEXICAL_BENCH_HEAP_MIB=4096`; the local defaults remain 16 workers and
8192 MiB. Each lifecycle requires a new output directory and retains its raw
results and logs in the `lexical-lookup-performance` artifact. Graph and persisted
call mismatches fail the job. The existing first-pair speed screen reports and
stops a non-improving run; it does not make one noisy timing pair a CI failure.
### Repeated baseline, 2026-11-03 (CI corpus)
The [baseline receipt](lexical-lookup/baseline-20261103.json) records three
independent sessions per lifecycle, each with three paired edits. Every pair
was retained, including slower candidates. Runtime order was reversed in the
middle session. The corpus is the sparse CI source tree at
`4916e0d5e1f4d6d70d41c1f39cd244bca2c7b390`; baseline and candidate runtimes are
`4916e0d5e1f4d6d70d41c1f39cd244bca2c7b390` and
`6b487a529e14d59e51a9df9a44524e98705a6b71`. Both use Node 22.23.3,
LadybugDB 0.22.0, two workers and a 4 GiB heap on the same shared host.
These are medians of the three session medians; cold seeds are excluded:
| Lifecycle | Baseline | PR candidate | Change |
| --- | ---: | ---: | ---: |
| Separate processes | 28.494s | 27.552s | 3.31% less time |
| Repeated processes | 30.276s | 29.814s | 1.53% less time |
The candidate's session medians have coefficients of variation of **5.02%**
and **17.01%**, respectively. These small aggregate gains are inconclusive on
this host. The receipt also retains every paired percentage change, graph hash,
persisted call target and host-load reading. All 18 edit pairs matched 23,641
source/semantic nodes and 64,292 relationships, and the seeded differential
check passed 72,000 comparisons per session. Selection and reporting share the
same corpus; no holdout was used. This does not establish a full-repository or
watch-mode speedup.
During measurement, the PR advanced to
`6aeeace8e96603fb062dcc709e834042d5a6df74` by merging a Kotlin grammar update.
The lookup implementations and lockfile are unchanged, but the newer native
grammar and revision were not measured in that baseline. This receipt remains
pinned to the earlier builds.
The subsequent [optimization report](../../docs/optimize/lexical-3805-report.md)
and [experiment receipt](lexical-lookup/optimization-20261103.json) compare two
additional candidates against unchanged `6aeeace8` controls. Neither met the
required repeatable 5% warm-process improvement with at most 3% fresh-process
regression. The runtime remains unchanged; the retained changes make complete,
opposite-order baseline collection reproducible.
---
## Current measurements — 2026-10-31
### Environment and timing boundaries

View file

@ -84,3 +84,46 @@ An in-place verification failure keeps a non-FTS dirty phase: an FTS-only
repair cannot clear the marker and recertify the inconsistent graph.
The same protection applies when FTS returns but analyzer finalization fails
before the final identity scan: recovery still rebuilds the graph.
## Memory scaling
`memory.mjs` measures the compiled production reader in a fresh process, without
forcing GC between reads. It records full identity verification, elapsed time,
sampled heap/RSS, the process RSS high-water mark, and initial/final retained heap.
The native pool is 512 MiB for measurements and 1 GiB for fixture creation.
Build the runtime with `npm run build`. Save this configuration outside the repo:
```json
{
"runtime": "/absolute/path/to/GitNexus",
"storage": "/tmp/gnx-identity-memory",
"output": "/tmp/gnx-identity-memory/seed.json",
"count": 1048576,
"width": 0,
"mode": "seed",
"reader": "stream",
"iterations": 40
}
```
Run from `gitnexus/`:
```sh
node --expose-gc --max-old-space-size=4096 bench/incremental-write-integrity/memory.mjs /tmp/config.json
```
Then change `mode` to `measure` and choose a new output file. Run each measurement
in a new process and use one process per database at a time. To compare the
original array-based implementation, set `runtime` to its separately built
checkout and `reader` to `baseline`; keep the same fixture and memory limits.
Vary `count` with separate fixtures to measure node scaling. `width` adds that
many ASCII bytes to each of ID, name and path, so it measures identity size
independently of node count. Repeat each point and hold the largest size out of
the fitted curve. A finite RSS fit is evidence over that range, not a guarantee
about every native allocator or the rest of the indexing pipeline.
The streaming verifier retains expected-node maps and one decoded row rather
than a full table of JavaScript rows. The expected maps grow linearly with graph
size; the native query result and the original graph still consume memory.

View file

@ -66,9 +66,9 @@ const source = (file) => import(pathToFileURL(join(ROOT, file)).href);
await adapter.tryFlushWAL();
const samples = [];
let calls = 0;
const query = async (sql) => {
const query = async (sql, onRow) => {
calls++;
return adapter.executeQuery(sql);
await adapter.forEachQueryRow(sql, onRow);
};
for (let run = 0; run < 9; run++) {
calls = 0;
@ -80,17 +80,18 @@ const source = (file) => import(pathToFileURL(join(ROOT, file)).href);
if (run >= 2) samples.push(elapsed);
}
const first = graph.iterNodes().next().value;
const alteredQuery = async (sql) => {
const rows = await adapter.executeQuery(sql);
if (sql.includes('(n:`Function`)')) rows.find((r) => r.id === first.id).startLine++;
return rows;
const alteredQuery = async (sql, onRow) => {
await adapter.forEachQueryRow(sql, (row) => {
if (sql.includes('(n:`Function`)') && row.id === first.id) row.startLine++;
onRow(row);
});
};
await assert.rejects(
reconcileGraphNodeIdentities(graph, alteredQuery, 'negative-control'),
/startLine/,
);
await assert.rejects(
reconcileGraphNodeIdentities(graph, async () => [], 'negative-control'),
reconcileGraphNodeIdentities(graph, async () => {}, 'negative-control'),
/missing ID/,
);
@ -124,12 +125,12 @@ const source = (file) => import(pathToFileURL(join(ROOT, file)).href);
rows.length !== size.nodes || mismatched > 0 || actual.size !== wanted.size;
if (rejected) {
await assert.rejects(
reconcileGraphNodeIdentities(graph, adapter.executeQuery, phase),
reconcileGraphNodeIdentities(graph, adapter.forEachQueryRow, phase),
/Graph identity reconciliation/,
);
} else {
assert.equal(
(await reconcileGraphNodeIdentities(graph, adapter.executeQuery, phase)).nodes,
(await reconcileGraphNodeIdentities(graph, adapter.forEachQueryRow, phase)).nodes,
size.nodes,
);
}

View file

@ -0,0 +1,143 @@
// Build the runtime, then: node --expose-gc --max-old-space-size=4096 memory.mjs CONFIG.json
// Config: { runtime, storage, output, count, width, mode: "seed"|"measure",
// reader: "baseline"|"stream", iterations }. Each measurement is a fresh process.
// No forced GC between reads. Initial/final GC measure retained graph memory only.
import fs from 'node:fs';
import path from 'node:path';
import assert from 'node:assert/strict';
import { createHash } from 'node:crypto';
import { pathToFileURL } from 'node:url';
const config = JSON.parse(fs.readFileSync(process.argv[2], 'utf8'));
const { runtime, storage, output, count, width = 0, mode, reader, iterations } = config;
assert(Number.isSafeInteger(count) && count >= 0);
assert(Number.isSafeInteger(width) && width >= 0);
assert(Number.isSafeInteger(iterations) && iterations > 0);
assert(['seed', 'measure'].includes(mode));
assert(['baseline', 'stream'].includes(reader));
assert(global.gc, 'Run with --expose-gc');
const bufferPoolBytes = (mode === 'seed' ? 1024 : 512) * 1048576;
process.env.GITNEXUS_LBUG_BUFFER_POOL_SIZE = String(bufferPoolBytes);
process.env.GITNEXUS_SHARED_STORE = 'off';
process.env.GITNEXUS_HOME = path.join(path.dirname(output), 'home');
const source = (file) => import(pathToFileURL(path.join(runtime, 'gitnexus/dist', file)).href);
const { createKnowledgeGraph } = await source('core/graph/graph.js');
const { reconcileGraphNodeIdentities } = await source('core/incremental/write-reconciliation.js');
const adapter = await source('core/lbug/lbug-adapter.js');
const graph = createKnowledgeGraph();
const padding = 'x'.repeat(width);
const fingerprint = createHash('sha256');
let identityBytes = 0;
for (let i = 0; i < count; i++) {
const index = String(i).padStart(7, '0');
const id = `Function:node${index}${padding}`;
const name = `fn${index}${padding}`;
const filePath = `src/file${index}.ts${padding}`;
const startLine = (i % 32) * 4;
const endLine = startLine + 2;
identityBytes += Buffer.byteLength(id) + Buffer.byteLength(name) + Buffer.byteLength(filePath);
fingerprint.update(JSON.stringify([id, name, filePath, startLine, endLine]) + '\n');
graph.addNode({ id, label: 'Function', properties: { name, filePath, startLine, endLine } });
}
const result = {
config,
node: process.version,
driver: JSON.parse(
fs.readFileSync(path.join(runtime, 'gitnexus/node_modules/@ladybugdb/core/package.json')),
).version,
bufferPoolBytes,
identityBytes,
identityFingerprint: fingerprint.digest('hex'),
iterations: [],
completed: false,
};
let peakRss = 0,
peakHeap = 0,
currentPeakRss = 0,
currentPeakHeap = 0;
function sample() {
const memory = process.memoryUsage();
peakRss = Math.max(peakRss, memory.rss);
peakHeap = Math.max(peakHeap, memory.heapUsed);
currentPeakRss = Math.max(currentPeakRss, memory.rss);
currentPeakHeap = Math.max(currentPeakHeap, memory.heapUsed);
return memory;
}
function save() {
result.peakRssBytes = Math.max(peakRss, process.resourceUsage().maxRSS * 1024);
result.sampledPeakHeapBytes = peakHeap;
fs.writeFileSync(output, JSON.stringify(result, null, 2) + '\n');
}
let timer;
try {
await adapter.initLbug(path.join(storage, 'lbug'), { skipFts: true });
if (mode === 'seed') {
await adapter.loadGraphToLbug(graph, storage, storage, undefined, undefined, undefined, 'none');
await adapter.tryFlushWAL();
} else {
global.gc();
result.initial = sample();
timer = setInterval(sample, 20);
timer.unref();
for (let i = 0; i < iterations; i++) {
currentPeakRss = 0;
currentPeakHeap = 0;
const before = sample();
let queries = 0;
const readRows =
reader === 'baseline'
? async (query) => {
queries++;
sample();
const rows = await adapter.executeQuery(query);
sample();
return rows;
}
: async (query, visit) => {
queries++;
sample();
await adapter.forEachQueryRow(query, visit);
sample();
};
const start = performance.now();
const receipt = await reconcileGraphNodeIdentities(graph, readRows, 'memory-scaling');
const ms = performance.now() - start;
assert.equal(receipt.nodes, count);
assert.equal(queries, receipt.tables);
const after = sample();
result.iterations.push({
i,
ms,
queries,
verifiedNodes: receipt.nodes,
before,
after,
peakRssBytes: currentPeakRss,
sampledPeakHeapBytes: currentPeakHeap,
maxRssKiB: process.resourceUsage().maxRSS,
});
save();
}
clearInterval(timer);
global.gc();
result.finalAfterGc = sample();
}
result.completed = true;
} catch (error) {
result.error = String(error);
process.exitCode = 1;
} finally {
clearInterval(timer);
await adapter.closeLbug();
save();
}
console.log(
JSON.stringify({
count,
width,
reader,
completed: result.completed,
iterations: result.iterations.length,
peakRssMiB: result.peakRssBytes / 1048576,
}),
);

File diff suppressed because it is too large Load diff

View file

@ -0,0 +1,109 @@
import assert from 'node:assert/strict';
import path from 'node:path';
import { pathToFileURL } from 'node:url';
const [baseline, candidate] = process.argv.slice(2);
assert(baseline && candidate, 'Expected BASE_RUNTIME CANDIDATE_RUNTIME');
const moduleAt = (root) =>
pathToFileURL(path.resolve(root, 'gitnexus/dist/_shared/scope-resolution/name-claims.js')).href;
const { lookupLexicalName: before } = await import(moduleAt(baseline));
const { lookupLexicalName: after } = await import(moduleAt(candidate));
let seed = 48161;
const random = (n) => {
seed = (Math.imul(seed, 1664525) + 1013904223) >>> 0;
return seed % n;
};
const pos = (n) => ({ startLine: n, startCol: 0 });
const range = (n) => ({ ...pos(n), endLine: n, endCol: 20 });
const names = ['a', 'b', 'missing'];
let comparisons = 0;
for (let trial = 0; trial < 3000; trial++) {
const scopes = new Map();
for (let i = 0; i < 5; i++) {
const claims = [];
const bindings = new Map();
const types = new Map();
const imports = [];
for (const name of names.slice(0, 2)) {
if (random(3) === 0)
claims.push({
name,
kind: ['binding', 'import', 'blocked'][random(3)],
range: range(2 + random(8)),
purpose: ['value', 'type', 'both', undefined][random(4)],
availableFrom: random(2) ? pos(2 + random(8)) : undefined,
inactive: ['outer', 'blocked', 'module'][random(3)],
redirect: random(8) === 0 ? 'module' : undefined,
hoisted: random(2) === 0,
merge: random(3) === 0,
});
if (random(2))
bindings.set(name, [
{
def: { nodeId: `${i}:${name}`, type: 'Function', filePath: 'a.ts' },
origin: random(2) ? 'local' : 'import',
availableFrom: random(2) ? pos(5) : undefined,
declarationRange: range(5),
},
]);
if (random(3) === 0)
imports.push({
kind: 'namespace',
localName: name,
targetFile: random(2) ? 'b.ts' : null,
targetExportedName: '',
atRange: range(4),
linkStatus: random(2) ? 'resolved' : 'unresolved',
});
if (random(3) === 0)
types.set(name, {
rawName: 'Value',
source: [
'annotation',
'constructor-inferred',
'assignment-inferred',
'receiver-propagated',
][random(4)],
declaredAtScope: String(i),
bindingRange: random(2) ? range(5) : undefined,
});
}
scopes.set(String(i), {
id: String(i),
parent: i === 0 ? null : String(i - 1),
kind: i === 0 ? 'Module' : ['Function', 'Block', 'Class', 'Object'][random(4)],
filePath: 'a.ts',
range: range(i),
bindings,
imports,
ownedDefs: [],
typeBindings: types,
nameClaims: claims,
lexicalNames: random(3) === 0 ? new Set(['a']) : undefined,
ownsReceivers: random(3) === 0 ? new Set(['b']) : undefined,
lookupPolicy: {
nameClaimsComplete: random(2) === 0,
deferParentActivation: random(2) === 0,
skipFromChildren: random(4) === 0,
},
});
}
if (trial % 20 === 0) scopes.get('0').parent = '4';
const sources = { scopes: { getScope: (id) => scopes.get(id) } };
for (const name of names)
for (const purpose of ['value', 'type'])
for (const position of [undefined, pos(1), pos(6), pos(12)]) {
const options = Object.freeze({
purpose,
position,
skipEnclosingClasses: trial % 2 === 0,
preservePosition: trial % 3 === 0,
});
assert.deepEqual(
after('4', name, sources, options),
before('4', name, sources, options),
`Trial ${trial}, ${name}, ${purpose}, line ${position?.startLine}`,
);
comparisons++;
}
}
console.log({ comparisons, status: 'passed' });

View file

@ -0,0 +1,303 @@
// Build both runtimes, then run with Node:
// measure.mjs BASE_RUNTIME CANDIDATE_RUNTIME BASE_CORPUS CANDIDATE_CORPUS OUTPUT [fresh|persistent]
// Corpus checkouts must be clean/disposable and at the same commit. OUTPUT must not exist.
import assert from 'node:assert/strict';
import fs from 'node:fs';
import path from 'node:path';
import os from 'node:os';
import { createHash } from 'node:crypto';
import { execFileSync, fork } from 'node:child_process';
import { fileURLToPath } from 'node:url';
const [baselineRuntime, candidateRuntime, baseline, candidate, output] = process.argv
.slice(2, 7)
.map((value) => path.resolve(value));
const lifecycle = process.argv[7] ?? 'fresh';
const workers = Number(process.env.GITNEXUS_LEXICAL_BENCH_WORKERS ?? 16);
const heapMiB = Number(process.env.GITNEXUS_LEXICAL_BENCH_HEAP_MIB ?? 8192);
const screenSetting = process.env.GITNEXUS_LEXICAL_BENCH_SCREEN ?? '1';
const baseFirstSetting = process.env.GITNEXUS_LEXICAL_BENCH_BASE_FIRST ?? '1';
assert(Number.isSafeInteger(workers) && workers > 0, 'Workers must be a positive integer');
assert(Number.isSafeInteger(heapMiB) && heapMiB > 0, 'Heap MiB must be a positive integer');
assert(['0', '1'].includes(screenSetting), 'Screen must be 0 or 1');
assert(['0', '1'].includes(baseFirstSetting), 'Base first must be 0 or 1');
const screen = screenSetting === '1';
const baseFirst = baseFirstSetting === '1';
assert(['fresh', 'persistent'].includes(lifecycle));
assert(
baseline && candidate && output,
'Expected BASE_RUNTIME CANDIDATE_RUNTIME BASE_CORPUS CANDIDATE_CORPUS OUTPUT [fresh|persistent]',
);
assert.notEqual(baseline, candidate);
const isWorker = process.env.GITNEXUS_LEXICAL_BENCH_WORKER === '1';
assert(baselineRuntime && candidateRuntime);
let runFullAnalysis, initLbug, executeQuery, closeLbug;
if (!isWorker) assert(!fs.existsSync(output), 'Output already exists');
const control = 'gitnexus/src/semantic-bench-control.ts';
const source =
'export class SemanticBenchStore { save(): void {} load(): void {} }\nexport function semanticBenchUse(s: SemanticBenchStore) { s.save(); }\n';
const head = (repo) =>
execFileSync('git', ['rev-parse', 'HEAD'], { cwd: repo, encoding: 'utf8' }).trim();
assert.equal(head(baseline), head(candidate));
for (const repo of isWorker ? [] : [baseline, candidate]) {
assert.equal(
execFileSync('git', ['status', '--porcelain'], { cwd: repo, encoding: 'utf8' }),
'',
'Checkout must be clean',
);
assert(!fs.existsSync(path.join(repo, control)));
}
fs.mkdirSync(output, { recursive: true });
process.env.GITNEXUS_HOME = path.join(output, 'home');
process.env.GITNEXUS_STORAGE_ROOT = path.join(output, 'indexes');
delete process.env.GITNEXUS_STORAGE_PATH;
process.env.GITNEXUS_SHARED_STORE = 'off';
process.env.PROF_SCOPE_RESOLUTION = '1';
const records = [];
if (isWorker) {
const runtime = process.env.GITNEXUS_BENCH_RUNTIME;
({ runFullAnalysis } = await import(path.join(runtime, 'gitnexus/dist/core/run-analyze.js')));
({ initLbug, executeQuery, closeLbug } = await import(
path.join(runtime, 'gitnexus/dist/core/lbug/lbug-adapter.js')
));
const { configureLoggerSink } = await import(path.join(runtime, 'gitnexus/dist/core/logger.js'));
configureLoggerSink({
destination: {
write(line) {
try {
records.push(JSON.parse(line));
} catch {
/* non-JSON diagnostic */
}
process.stderr.write(line);
},
},
});
}
const results = [];
const canonical = (value) =>
Array.isArray(value)
? value.map(canonical)
: value && typeof value === 'object'
? Object.fromEntries(
Object.keys(value)
.sort()
.map((key) => [key, canonical(value[key])]),
)
: value;
const digest = (rows) => {
const hash = createHash('sha256');
for (const row of rows.sort()) hash.update(row).update('\n');
return { count: rows.length, sha256: hash.digest('hex') };
};
async function sample(repo, reuse, name) {
process.env.GITNEXUS_EXPERIMENTAL_SEMANTIC_CACHE = '0';
global.gc?.();
records.length = 0;
const logs = [];
const progress = [];
const loadBefore = os.loadavg();
const started = performance.now();
const result = await runFullAnalysis(
repo,
{
skipAgentsMd: true,
skipSkills: true,
workerPoolSize: workers,
verbose: true,
},
{
onProgress(phase, percent, message) {
if (progress.at(-1)?.phase !== phase)
progress.push({ ms: performance.now() - started, phase, percent, message });
},
onLog(message) {
logs.push(message);
},
},
);
const ms = performance.now() - started;
assert(!result.alreadyUpToDate, 'A no-op is not an edit benchmark');
assert(!result.graphWriteCollapsed);
assert(!result.ftsSkipped, 'This benchmark includes FTS');
const graph = result.pipelineResult.graph;
const nodes = [],
edges = [];
for (const node of graph.iterNodes()) {
// Derived clustering is stochastic; compare all source/semantic properties.
if (['Community', 'Process', 'Folder'].includes(node.label)) continue;
nodes.push(JSON.stringify(canonical(node)));
}
let controlCalls = 0;
for (const edge of graph.iterRelationships()) {
if (['MEMBER_OF', 'STEP_IN_PROCESS', 'CONTAINS'].includes(edge.type)) continue;
edges.push(JSON.stringify(canonical(edge)));
if (edge.type === 'CALLS' && edge.sourceId.includes('semanticBenchUse')) controlCalls++;
}
assert(controlCalls > 0, 'The changed call must resolve');
const expectedTarget = fs.readFileSync(path.join(repo, control), 'utf8').includes('s.load()')
? 'load'
: 'save';
let persistedTargets;
try {
await initLbug(path.join(result.storagePath, 'lbug'), { skipFts: true });
persistedTargets = await executeQuery(
"MATCH (a:Function)-[r:CodeRelation]->(b:Method) WHERE a.name = 'semanticBenchUse' AND r.type = 'CALLS' RETURN b.name AS target",
);
assert.deepEqual(
persistedTargets.map((row) => row.target),
[expectedTarget],
'Persisted call must follow the edit',
);
} finally {
await closeLbug();
}
const report = {
name,
variant: reuse ? 'candidate' : 'baseline',
ms,
loadBefore,
loadAfter: os.loadavg(),
maxRssKiB: process.resourceUsage().maxRSS,
stats: result.stats,
incremental: result.incrementalStats,
rebuildReasons: result.rebuildReasons,
reparsedFiles: result.pipelineResult.reparsedFileCount,
cachedFiles: result.pipelineResult.parseCacheHitFileCount,
graph: { nodes: digest(nodes), edges: digest(edges) },
logs,
progress,
persistedTargets,
records: [...records],
};
console.log(
'LEXICAL_BENCH ' +
JSON.stringify({ name, ms, incremental: report.incremental, graph: report.graph }),
);
return report;
}
// Separate runtimes isolate implementation and JIT state. Fresh mode exits
// after every sample; persistent mode retains a process across real edits.
function worker(repo, reuse) {
let child, pending, exited;
const start = () => {
child = fork(fileURLToPath(import.meta.url), process.argv.slice(2), {
execArgv: [`--max-old-space-size=${heapMiB}`, '--expose-gc'],
env: {
...process.env,
GITNEXUS_LEXICAL_BENCH_WORKER: '1',
GITNEXUS_BENCH_RUNTIME: reuse ? candidateRuntime : baselineRuntime,
},
stdio: ['ignore', 'inherit', 'inherit', 'ipc'],
});
exited = new Promise((resolve) => child.once('exit', resolve));
child.on('message', (message) => {
const current = pending;
pending = undefined;
if (message.error) current?.reject(new Error(message.error));
else current?.resolve(message.report);
});
child.on('error', (error) => pending?.reject(error));
child.on('exit', (code) => pending?.reject(new Error(`Benchmark worker exited: ${code}`)));
};
const close = async () => {
if (!child) return;
if (child.connected) child.disconnect();
const code = await exited;
child = undefined;
assert.equal(code, 0, 'Worker must exit normally after analysis');
};
return {
async run(name) {
assert(!pending, 'Samples must run serially');
if (!child) start();
const report = await new Promise((resolve, reject) => {
pending = { resolve, reject };
child.send({ repo, reuse, name });
});
results.push(report);
fs.writeFileSync(
path.join(output, 'results.json'),
JSON.stringify(
{
head: head(repo),
node: process.version,
lifecycle,
workers,
heapMiB,
screen,
baseFirst,
baselineRuntime,
candidateRuntime,
results,
},
null,
2,
),
);
if (lifecycle === 'fresh') await close();
return report;
},
close,
};
}
if (isWorker) {
process.on('message', async ({ repo, reuse, name }) => {
try {
process.send({ report: await sample(repo, reuse, name) });
} catch (error) {
process.send({ error: error.stack ?? String(error) });
}
});
} else {
const beforeWorker = worker(baseline, false);
const afterWorker = worker(candidate, true);
try {
for (const repo of [baseline, candidate]) {
fs.mkdirSync(path.dirname(path.join(repo, control)), { recursive: true });
fs.writeFileSync(path.join(repo, control), source);
}
let beforeSeed, afterSeed;
if (baseFirst) {
beforeSeed = await beforeWorker.run('baseline-seed');
afterSeed = await afterWorker.run('candidate-seed');
} else {
afterSeed = await afterWorker.run('candidate-seed');
beforeSeed = await beforeWorker.run('baseline-seed');
}
assert.deepEqual(afterSeed.graph, beforeSeed.graph, 'Seed semantic graphs differ');
for (let pair = 1; pair <= 3; pair++) {
const text = pair % 2 ? source.replace('s.save()', 's.load()') : source;
for (const repo of [baseline, candidate]) fs.writeFileSync(path.join(repo, control), text);
// Alternate ordering to reduce warm-up/load bias. Both indexes see a real edit.
let before, after;
if (Boolean(pair % 2) === baseFirst) {
before = await beforeWorker.run(`baseline-${pair}`);
after = await afterWorker.run(`candidate-${pair}`);
} else {
after = await afterWorker.run(`candidate-${pair}`);
before = await beforeWorker.run(`baseline-${pair}`);
}
assert.deepEqual(
after.graph,
before.graph,
'Semantic graph differs from a full semantic pass',
);
for (const entry of [before, after]) {
assert.equal(entry.incremental?.writeMode, 'incremental');
assert.equal(entry.incremental?.changedFiles, 1);
assert.deepEqual(entry.rebuildReasons, []);
}
if (screen && pair === 1 && after.ms >= before.ms) {
console.log('SCREEN REJECTED: no end-to-end improvement; stopping after one pair.');
break;
}
}
} finally {
const closed = await Promise.allSettled([beforeWorker.close(), afterWorker.close()]);
for (const repo of [baseline, candidate]) fs.rmSync(path.join(repo, control), { force: true });
for (const result of closed) if (result.status === 'rejected') throw result.reason;
}
}

View file

@ -0,0 +1,735 @@
{
"date": "2026-11-02",
"baseCommit": "347cc99b4a237f9a71c0084167bcbcbd9d06e610",
"node": "v22.23.3",
"ladybug": "0.22.0",
"hardware": {
"cpu": "12th Gen Intel(R) Core(TM) i9-12900K",
"logicalCpus": 24,
"platform": "linux",
"arch": "x64"
},
"settings": {
"workers": 16,
"heapMiB": 8192,
"sharedHost": true,
"scopePhaseLogging": true,
"cpuProfile": false,
"independentColdSeeds": true
},
"measurement": {
"timed": "runFullAnalysis including graph persistence and FTS",
"excluded": ["process startup", "graph hashing", "post-analysis persisted target query"],
"editOrder": [
"baseline-1",
"candidate-1",
"candidate-2",
"baseline-2",
"baseline-3",
"candidate-3"
],
"graphExclusions": [
"Folder",
"Community",
"Process",
"CONTAINS",
"MEMBER_OF",
"STEP_IN_PROCESS"
],
"limitations": [
"three fresh-process pairs and one repeated-process pair on a shared host",
"graph hashes are post-persistence; source content may have been cleared",
"persistent process exercises repeated analysis, not watch event scheduling",
"baseline resolution limits also apply to candidate",
"process maxRSS includes earlier samples in persistent mode"
]
},
"pin": {
"patchSha256": "a61b37f5ab42a95a62ae96510a5e285cb9b835a3f983ce4a96ecd0413504ea43",
"runtimes": {
"baseline": {
"gitnexus/dist/_shared/scope-resolution/name-claims.js": "157fd1ff2870468d335f97fff41d0b002f4cad749db64634aa48469053d7cea3",
"gitnexus/dist/core/ingestion/scope-resolution/scope/walkers.js": "15dba69d6ec871e43f4d0d1aadbad93e1645c13a9ad2e664e9f9bd679f73e55c"
},
"candidate": {
"gitnexus/dist/_shared/scope-resolution/name-claims.js": "6f998066bcb3f17d69f2dbae7dba1a1d5af865bc8798d4b137aaed9387a81284",
"gitnexus/dist/core/ingestion/scope-resolution/scope/walkers.js": "9400ed640a22343f6cc54e3515944af664022ac7f4a8c4306da0cc95df37f9e1"
}
}
},
"harnessSha256": "dbd428561cd614b8f0fb2dda48ad5445e214bea02a9c39f201cc2488bd7a81c8",
"harnessProvenance": "Checksums identify the scripts used for measurement, before repository formatting, preflight/cleanup fixes and configurable CI worker/heap limits. Local resource defaults are unchanged.",
"differential": {
"seed": 48161,
"comparisons": 72000,
"status": "passed",
"harnessSha256": "8a34834c153531daa9f6000e178fc87458e48d3507c8c6e20e2738f739413e88"
},
"runs": {
"fresh": {
"rawResultSha256": "3da8095b17ecbb5099164220152f8c9582268868db6a7d23b2565d987f2a7d6d",
"samples": [
{
"name": "baseline-seed",
"variant": "baseline",
"ms": 218218.857255,
"loadBefore": [19.08, 16.18, 14.03],
"loadAfter": [13.58, 16.53, 14.82],
"maxRssKiB": 5929416,
"rebuildReasons": [],
"reparsedFiles": 3120,
"cachedFiles": 0,
"phaseMs": {
"extracting": 307.5897699999996,
"structure": 59.0564260000001,
"parsing": 77980.037363,
"scopeResolution": 80390.527192,
"enriching": 36.09870200001751,
"communities": 2342.1585379999888,
"processes": 659.8627029999916,
"complete": 0.06783200000063516,
"lbug": 17391.16107600002,
"fts": 29286.008015999978,
"done": 6083.309720000019,
"beforeFirstProgress": 3682.979917
},
"scopeLogs": [
"[scope-resolution prof] extract=21ms finalize=71ms propagate=78ms rangeBind=0.0ms resolve=175ms emit=1547ms total=1891ms (110 files)",
"[scope-resolution prof] extract=2649ms finalize=1451ms propagate=1426ms rangeBind=0.0ms resolve=7468ms emit=41291ms total=54285ms (2893 files)",
"[scope-resolution prof] extract=270ms finalize=65ms propagate=61ms rangeBind=0.0ms resolve=444ms emit=1835ms total=2675ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "candidate-seed",
"variant": "candidate",
"ms": 187147.92687899998,
"loadBefore": [13.13, 16.39, 14.78],
"loadAfter": [11.64, 15.07, 14.65],
"maxRssKiB": 5912680,
"rebuildReasons": [],
"reparsedFiles": 3120,
"cachedFiles": 0,
"phaseMs": {
"extracting": 248.5604829999993,
"structure": 55.20889800000077,
"parsing": 62435.270292,
"scopeResolution": 61285.705398999984,
"enriching": 31.93969999998808,
"communities": 2074.6883160000143,
"processes": 544.6516690000135,
"complete": 0.05536599998595193,
"lbug": 16684.85912899999,
"fts": 29567.160881000018,
"done": 5673.711943999981,
"beforeFirstProgress": 8546.114802
},
"scopeLogs": [
"[scope-resolution prof] extract=18ms finalize=60ms propagate=52ms rangeBind=0.0ms resolve=118ms emit=1176ms total=1425ms (110 files)",
"[scope-resolution prof] extract=2048ms finalize=1153ms propagate=1045ms rangeBind=0.0ms resolve=5119ms emit=31736ms total=41101ms (2893 files)",
"[scope-resolution prof] extract=201ms finalize=62ms propagate=64ms rangeBind=0.0ms resolve=282ms emit=1469ms total=2078ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "baseline-1",
"variant": "baseline",
"ms": 122065.065072,
"loadBefore": [11.43, 14.97, 14.62],
"loadAfter": [8.47, 12.61, 13.79],
"maxRssKiB": 4474272,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 221.9082469999994,
"structure": 52.67533299999923,
"parsing": 8765.232970000001,
"scopeResolution": 65945.25450899999,
"enriching": 121.54619799999637,
"complete": 0.25131200000760145,
"lbug": 10923.801725000012,
"communities": 2019.391571,
"processes": 5888.645534999989,
"fts": 13240.190296,
"done": 8270.738006,
"beforeFirstProgress": 6615.429370000001
},
"scopeLogs": [
"[scope-resolution prof] extract=19ms finalize=60ms propagate=58ms rangeBind=0.0ms resolve=148ms emit=1294ms total=1578ms (110 files)",
"[scope-resolution prof] extract=2031ms finalize=1178ms propagate=1108ms rangeBind=0.0ms resolve=6717ms emit=35251ms total=46284ms (2893 files)",
"[scope-resolution prof] extract=232ms finalize=61ms propagate=59ms rangeBind=0.0ms resolve=379ms emit=1552ms total=2283ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
},
{
"name": "candidate-1",
"variant": "candidate",
"ms": 111134.852017,
"loadBefore": [8.47, 12.61, 13.79],
"loadAfter": [7.79, 11.1, 13.1],
"maxRssKiB": 4406608,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 228.39688900000056,
"structure": 58.946551,
"parsing": 7914.542683,
"scopeResolution": 62759.170073999994,
"enriching": 126.18967100000009,
"complete": 0.17001499999605585,
"lbug": 9233.545855000018,
"communities": 1969.2113839999947,
"processes": 1475.3109669999976,
"fts": 12857.877876999992,
"done": 8347.120962,
"beforeFirstProgress": 6164.369089
},
"scopeLogs": [
"[scope-resolution prof] extract=15ms finalize=57ms propagate=51ms rangeBind=0.0ms resolve=115ms emit=1221ms total=1459ms (110 files)",
"[scope-resolution prof] extract=2192ms finalize=1268ms propagate=1122ms rangeBind=0.0ms resolve=5371ms emit=33985ms total=43937ms (2893 files)",
"[scope-resolution prof] extract=176ms finalize=75ms propagate=55ms rangeBind=0.0ms resolve=253ms emit=1333ms total=1892ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
},
{
"name": "candidate-2",
"variant": "candidate",
"ms": 101730.29475,
"loadBefore": [7.79, 11.1, 13.1],
"loadAfter": [8.7, 10.56, 12.7],
"maxRssKiB": 4447528,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 227.87885000000006,
"structure": 50.07236400000011,
"parsing": 6870.926647,
"scopeResolution": 58457.57910199999,
"enriching": 123.74990100000286,
"complete": 0.12248400000680704,
"lbug": 9597.999660000001,
"communities": 1855.2971499999985,
"processes": 1319.511560999992,
"fts": 12526.206298000005,
"done": 8028.759919000004,
"beforeFirstProgress": 2672.190814
},
"scopeLogs": [
"[scope-resolution prof] extract=14ms finalize=61ms propagate=47ms rangeBind=0.0ms resolve=103ms emit=1077ms total=1302ms (110 files)",
"[scope-resolution prof] extract=1860ms finalize=1052ms propagate=1064ms rangeBind=0.0ms resolve=4989ms emit=31810ms total=40774ms (2893 files)",
"[scope-resolution prof] extract=189ms finalize=62ms propagate=52ms rangeBind=0.0ms resolve=258ms emit=1383ms total=1943ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "baseline-2",
"variant": "baseline",
"ms": 114005.953714,
"loadBefore": [8.7, 10.56, 12.7],
"loadAfter": [9.5, 10.18, 12.3],
"maxRssKiB": 4501784,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 220.18439399999988,
"structure": 56.05602999999974,
"parsing": 6693.682359,
"scopeResolution": 68196.749623,
"enriching": 136.36504800000694,
"complete": 0.24019299999054056,
"lbug": 10199.119456000015,
"communities": 1921.6023549999954,
"processes": 1269.8549959999946,
"fts": 12961.042190000007,
"done": 8423.348173999999,
"beforeFirstProgress": 3927.708896
},
"scopeLogs": [
"[scope-resolution prof] extract=15ms finalize=63ms propagate=59ms rangeBind=0.0ms resolve=129ms emit=1239ms total=1506ms (110 files)",
"[scope-resolution prof] extract=2317ms finalize=1312ms propagate=1295ms rangeBind=0.0ms resolve=7098ms emit=36604ms total=48627ms (2893 files)",
"[scope-resolution prof] extract=225ms finalize=58ms propagate=59ms rangeBind=0.0ms resolve=366ms emit=1623ms total=2331ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "baseline-3",
"variant": "baseline",
"ms": 131285.117027,
"loadBefore": [9.94, 10.26, 12.32],
"loadAfter": [11.35, 10.88, 12.28],
"maxRssKiB": 4423784,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 246.84667200000013,
"structure": 55.95557000000008,
"parsing": 7195.49225,
"scopeResolution": 72694.88498100001,
"enriching": 136.15589299998828,
"complete": 0.1802390000084415,
"lbug": 13022.899677999987,
"communities": 2414.3188599999994,
"processes": 1639.1324740000127,
"fts": 14523.853445,
"done": 16550.80450099999,
"beforeFirstProgress": 2804.592464
},
"scopeLogs": [
"[scope-resolution prof] extract=17ms finalize=67ms propagate=66ms rangeBind=0.0ms resolve=138ms emit=1370ms total=1658ms (110 files)",
"[scope-resolution prof] extract=2113ms finalize=1144ms propagate=1129ms rangeBind=0.0ms resolve=7406ms emit=40471ms total=52262ms (2893 files)",
"[scope-resolution prof] extract=255ms finalize=74ms propagate=68ms rangeBind=0.0ms resolve=393ms emit=1821ms total=2611ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
},
{
"name": "candidate-3",
"variant": "candidate",
"ms": 111303.319634,
"loadBefore": [11, 10.82, 12.25],
"loadAfter": [8.98, 10.27, 11.9],
"maxRssKiB": 4458828,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 224.80108500000006,
"structure": 54.754461999999876,
"parsing": 9906.336681,
"scopeResolution": 62344.37333,
"enriching": 117.25856000000203,
"complete": 0.1237049999908777,
"lbug": 9196.625813000006,
"communities": 2015.0293479999964,
"processes": 3180.0593300000037,
"fts": 12870.898965,
"done": 7974.200893999994,
"beforeFirstProgress": 3418.8574609999996
},
"scopeLogs": [
"[scope-resolution prof] extract=22ms finalize=74ms propagate=61ms rangeBind=0.0ms resolve=115ms emit=1269ms total=1540ms (110 files)",
"[scope-resolution prof] extract=2053ms finalize=1125ms propagate=1129ms rangeBind=0.0ms resolve=5562ms emit=32778ms total=42647ms (2893 files)",
"[scope-resolution prof] extract=215ms finalize=61ms propagate=54ms rangeBind=0.0ms resolve=267ms emit=1445ms total=2041ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
}
],
"summary": {
"baseline": {
"editCount": 3,
"medianMs": 122065.065072,
"sumMs": 367356.13581300003,
"medianScopeMs": 68196.749623,
"maxRssKiB": 4501784
},
"candidate": {
"editCount": 3,
"medianMs": 111134.852017,
"sumMs": 324168.466401,
"medianScopeMs": 62344.37333,
"maxRssKiB": 4458828
},
"medianReductionPercent": 8.954415457488041,
"aggregateReductionPercent": 11.756349003514243,
"scopeMedianReductionPercent": 8.581605905490585
}
},
"persistent": {
"rawResultSha256": "b09e4f487c1933281a4f875620b5ac5a23794e56246d6c317dd6d3196762293e",
"samples": [
{
"name": "baseline-seed",
"variant": "baseline",
"ms": 158147.924033,
"loadBefore": [3.28, 6.72, 10.14],
"loadAfter": [5.3, 7.19, 9.87],
"maxRssKiB": 5757604,
"rebuildReasons": [],
"reparsedFiles": 3120,
"cachedFiles": 0,
"phaseMs": {
"extracting": 205.49587699999938,
"structure": 49.78513100000055,
"parsing": 58223.494267999995,
"scopeResolution": 62109.850408000006,
"enriching": 31.760779000003822,
"communities": 1872.6300760000013,
"processes": 506.77787499999977,
"complete": 0.04477599999518134,
"lbug": 10664.06607499998,
"fts": 18818.957672999997,
"done": 2247.716622000007,
"beforeFirstProgress": 3417.344473
},
"scopeLogs": [
"[scope-resolution prof] extract=16ms finalize=59ms propagate=66ms rangeBind=0.0ms resolve=132ms emit=1181ms total=1454ms (110 files)",
"[scope-resolution prof] extract=1988ms finalize=1096ms propagate=1099ms rangeBind=0.0ms resolve=6393ms emit=33758ms total=44335ms (2893 files)",
"[scope-resolution prof] extract=168ms finalize=64ms propagate=50ms rangeBind=0.0ms resolve=322ms emit=1375ms total=1978ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "candidate-seed",
"variant": "candidate",
"ms": 153951.283825,
"loadBefore": [5.3, 7.19, 9.87],
"loadAfter": [6.35, 7.83, 9.77],
"maxRssKiB": 5734340,
"rebuildReasons": [],
"reparsedFiles": 3120,
"cachedFiles": 0,
"phaseMs": {
"extracting": 226.53108299999963,
"structure": 46.397985999999946,
"parsing": 58254.662231,
"scopeResolution": 61421.754931999996,
"enriching": 35.70321300000069,
"communities": 1928.893007000006,
"processes": 529.0402800000011,
"complete": 0.05887399999483023,
"lbug": 10686.215123000002,
"fts": 16689.450012999994,
"done": 2236.283251999994,
"beforeFirstProgress": 1896.2938310000002
},
"scopeLogs": [
"[scope-resolution prof] extract=16ms finalize=54ms propagate=49ms rangeBind=0.0ms resolve=108ms emit=1134ms total=1361ms (110 files)",
"[scope-resolution prof] extract=2107ms finalize=1253ms propagate=1084ms rangeBind=0.0ms resolve=5107ms emit=32228ms total=41779ms (2893 files)",
"[scope-resolution prof] extract=172ms finalize=65ms propagate=52ms rangeBind=0.0ms resolve=270ms emit=1373ms total=1932ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "595afccf9f3fa2725727f3439cb5fd7b1aadd18ae6aca4eaa478ef00b793fe21"
}
},
"persistedTargets": [
{
"target": "save"
}
]
},
{
"name": "baseline-1",
"variant": "baseline",
"ms": 109482.844232,
"loadBefore": [6.35, 7.83, 9.77],
"loadAfter": [3.81, 6.46, 9.05],
"maxRssKiB": 5757604,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 211.5071590000298,
"structure": 41.07579499995336,
"parsing": 7984.799295000033,
"scopeResolution": 67012.942599,
"enriching": 124.94524899998214,
"complete": 0.18859700002940372,
"lbug": 8813.617234999954,
"communities": 1921.8943519999739,
"processes": 1619.4587620000239,
"fts": 10599.878043000004,
"done": 8947.475924000028,
"beforeFirstProgress": 2205.0612219999894
},
"scopeLogs": [
"[scope-resolution prof] extract=12ms finalize=49ms propagate=55ms rangeBind=0.0ms resolve=133ms emit=1216ms total=1467ms (110 files)",
"[scope-resolution prof] extract=2000ms finalize=1107ms propagate=1133ms rangeBind=0.0ms resolve=7186ms emit=37166ms total=48593ms (2893 files)",
"[scope-resolution prof] extract=164ms finalize=57ms propagate=58ms rangeBind=0.0ms resolve=346ms emit=1455ms total=2081ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
},
{
"name": "candidate-1",
"variant": "candidate",
"ms": 123364.35062099999,
"loadBefore": [3.81, 6.46, 9.05],
"loadAfter": [12.23, 8.31, 9.37],
"maxRssKiB": 5896624,
"incremental": {
"changedFiles": 1,
"reparsedFiles": 22,
"affectedDependents": 0,
"deletedFiles": 0,
"writeMode": "incremental"
},
"rebuildReasons": [],
"reparsedFiles": 22,
"cachedFiles": 3098,
"phaseMs": {
"extracting": 221.45905800000764,
"structure": 46.565340999979526,
"parsing": 7283.348788000003,
"scopeResolution": 65739.48159400001,
"enriching": 129.2759209999931,
"complete": 0.21377500001108274,
"lbug": 12901.81731199997,
"communities": 2457.193612000032,
"processes": 2556.050026999961,
"fts": 14765.58123000001,
"done": 14857.72775999998,
"beforeFirstProgress": 2405.636203000031
},
"scopeLogs": [
"[scope-resolution prof] extract=11ms finalize=47ms propagate=51ms rangeBind=0.0ms resolve=107ms emit=1124ms total=1340ms (110 files)",
"[scope-resolution prof] extract=2046ms finalize=1134ms propagate=1135ms rangeBind=0.0ms resolve=5161ms emit=34994ms total=44470ms (2893 files)",
"[scope-resolution prof] extract=407ms finalize=69ms propagate=71ms rangeBind=0.0ms resolve=369ms emit=2216ms total=3132ms (113 files)"
],
"graph": {
"nodes": {
"count": 114413,
"sha256": "d2e2ee693005fee4770a654b1391f5348d7b03316d57bf0e7bceb4cf6787f9ad"
},
"edges": {
"count": 318798,
"sha256": "4ece58529c66b824a32ee5b1c3f642193279f98e39d95eea53bf2f5ec1700ae8"
}
},
"persistedTargets": [
{
"target": "load"
}
]
}
],
"summary": {
"baseline": {
"editCount": 1,
"medianMs": 109482.844232,
"sumMs": 109482.844232,
"medianScopeMs": 67012.942599,
"maxRssKiB": 5757604
},
"candidate": {
"editCount": 1,
"medianMs": 123364.35062099999,
"sumMs": 123364.35062099999,
"medianScopeMs": 65739.48159400001,
"maxRssKiB": 5896624
},
"medianReductionPercent": -12.679161275335815,
"aggregateReductionPercent": -12.679161275335815,
"scopeMedianReductionPercent": 1.9003209762333162
}
}
},
"decision": "Experimental: fresh-process screen positive; persistent screen stopped after first non-improving pair; benefit for both workflows unproven",
"validation": {
"runtimePinsVerifiedAfterRuns": true,
"focusedTests": {
"files": 6,
"tests": 137,
"status": "passed"
},
"scopeTests": {
"files": 175,
"tests": 3467,
"status": "passed",
"overlapWithFocused": true
},
"differentialComparisons": 72000,
"parameterizedDifferentialRerun": "passed",
"build": "passed for both runtimes",
"testTypecheck": "passed",
"eslint": "passed; existing warnings",
"fullSuite": "not run"
}
}

File diff suppressed because it is too large Load diff

View file

@ -34,14 +34,15 @@ const storedIdentityField = (node: GraphNode, field: string): unknown => {
* retained rows. A write-set-only probe misses corruption of an unchanged row.
* Scan one label at a time, without path/range predicates: a corrupt field must
* not be able to hide its own row. No multi-label scans (#3139), N+1 ID lookups,
* source-content copies, or whole-graph row materialization.
* source-content copies, or JS row-array materialization. The reader calls the
* synchronous visitor for each row and releases its cursor before resolving.
*
* This certifies node identity, not relationships or native storage internals.
* Call after COPY/checkpoint and after FTS/embedding/checkpoint, before metadata.
*/
export async function reconcileGraphNodeIdentities(
graph: KnowledgeGraph,
query: (cypher: string) => Promise<IdentityRow[]>,
readRows: (cypher: string, onRow: (row: IdentityRow) => void) => Promise<void>,
phase: string,
): Promise<{ nodes: number; tables: number }> {
const expected = new Map<NodeTableName, Map<string, GraphNode>>();
@ -71,10 +72,25 @@ export async function reconcileGraphNodeIdentities(
{ cause },
);
};
let rows: IdentityRow[];
try {
rows = await query(
await readRows(
`MATCH (n:\`${table}\`) RETURN ${fields.map((f) => `n.${f} AS ${f}`).join(', ')}`,
(row) => {
const node = remaining.get(row.id);
if (!node) throw new Error(`unexpected or duplicate ID ${JSON.stringify(row.id)}`);
for (const field of identityFields) {
const wanted = storedIdentityField(node, field);
const actual =
field === 'name' || field === 'filePath' ? (row[field] ?? '') : row[field];
if (actual !== wanted) {
throw new Error(
`${row.id}.${field}: expected ${JSON.stringify(wanted)}, read ${JSON.stringify(actual)}`,
);
}
}
remaining.delete(row.id);
nodes++;
},
);
} catch (error) {
fail(
@ -82,21 +98,6 @@ export async function reconcileGraphNodeIdentities(
error,
);
}
for (const row of rows) {
const node = remaining.get(row.id);
if (!node) fail(`unexpected or duplicate ID ${JSON.stringify(row.id)}`);
for (const field of identityFields) {
const wanted = storedIdentityField(node, field);
const actual = field === 'name' || field === 'filePath' ? (row[field] ?? '') : row[field];
if (actual !== wanted) {
fail(
`${row.id}.${field}: expected ${JSON.stringify(wanted)}, read ${JSON.stringify(actual)}`,
);
}
}
remaining.delete(row.id);
nodes++;
}
if (remaining.size) {
fail(
`${remaining.size} missing ID(s), including ${JSON.stringify(remaining.keys().next().value)}`,

View file

@ -99,6 +99,7 @@ export function lookupNameClaim(
local,
scopes.bindings.get(scope.id)?.get(spelling) ?? EMPTY_BINDINGS,
) ?? EMPTY_BINDINGS;
if (additional.length === 0 && augmented.length === 0) return local;
return [
...local,
...compatibleClassBindings(additional, options.classCandidateFilter),
@ -108,24 +109,23 @@ export function lookupNameClaim(
),
];
}
return [
...local,
...compatibleClassBindings(
scopes.bindings.get(scope.id)?.get(spelling) ?? EMPTY_BINDINGS,
options.classCandidateFilter,
),
...augmented,
...(collectNamespaceFqnBindings(
scope.id,
spelling,
scopes,
options.classCandidateFilter,
) ?? []),
...compatibleClassBindings(
scopes.workspaceFqnBindings?.get(spelling) ?? EMPTY_BINDINGS,
options.classCandidateFilter,
),
];
const indexed = compatibleClassBindings(
scopes.bindings.get(scope.id)?.get(spelling) ?? EMPTY_BINDINGS,
options.classCandidateFilter,
);
const namespace =
collectNamespaceFqnBindings(scope.id, spelling, scopes, options.classCandidateFilter) ??
EMPTY_BINDINGS;
const workspace = compatibleClassBindings(
scopes.workspaceFqnBindings?.get(spelling) ?? EMPTY_BINDINGS,
options.classCandidateFilter,
);
if (
local.length + indexed.length + augmented.length + namespace.length + workspace.length ===
0
)
return EMPTY_BINDINGS;
return [...local, ...indexed, ...augmented, ...namespace, ...workspace];
},
},
options,

View file

@ -3,6 +3,7 @@ import { createReadStream, createWriteStream, existsSync, constants as fsConstan
import { createInterface } from 'readline';
import { once } from 'events';
import { finished } from 'stream/promises';
import { setImmediate as yieldQueryRows } from 'node:timers/promises';
import path from 'path';
import lbug from '@ladybugdb/core';
import { closeQueryResults } from './query-result-utils.js';
@ -1956,6 +1957,54 @@ export const executeQuery = async (cypher: string): Promise<any[]> => {
return await executePrepared(cypher, {});
};
/**
* Visit a single trusted, parameter-free query without retaining a JS row array.
* The visitor must be synchronous and must not issue another connection query.
* Keep the connection locked through yields and cursor close so a checkpoint
* cannot interleave with reading. Use direct execution: repeated prepared reads
* exhausted the native pool in the normal-GC identity scaling probe.
*/
export const forEachQueryRow = async (cypher: string, onRow: (row: any) => void): Promise<void> => {
warnIfQueryTextUnbounded(cypher, 'forEachQueryRow', (message) => logger.warn(message));
const c = conn;
if (!c) throw new Error('LadybugDB not initialized. Call initLbug first.');
await withConnLock(async () => {
const queryResult = await c.query(cypher);
const results = Array.isArray(queryResult) ? queryResult : [queryResult];
let failed = false;
let failure: unknown;
try {
if (results.length !== 1 || !results[0]) {
throw new Error('forEachQueryRow requires exactly one query result');
}
const result = results[0];
let rows = 0;
while (result.hasNext()) {
const row = result.getNextSync();
if (row === null) throw new Error('Query cursor ended before its advertised next row');
onRow(row);
if (++rows % 1024 === 0) await yieldQueryRows();
}
} catch (error) {
failed = true;
failure = error;
} finally {
for (const result of results) {
try {
result?.close();
} catch (error) {
if (!failed) {
failed = true;
failure = error;
}
}
}
}
if (failed) throw failure;
});
};
export const streamQuery = async (
cypher: string,
onRow: (row: any) => void | Promise<void>,

View file

@ -73,6 +73,7 @@ import {
loadGraphToLbug,
getLbugStats,
executeQuery,
forEachQueryRow,
executeWithReusedStatement,
closeLbug,
closeLbugBeforeExit,
@ -4419,7 +4420,7 @@ async function runFullAnalysisInner(
await markIncrementalGraphVerification();
await reconcileGraphNodeIdentities(
pipelineResult.graph,
executeQuery,
forEachQueryRow,
'post-COPY/checkpoint',
);
}
@ -5573,7 +5574,7 @@ async function runFullAnalysisInner(
}
await reconcileGraphNodeIdentities(
pipelineResult.graph,
executeQuery,
forEachQueryRow,
'pre-publish/checkpoint',
);
}

View file

@ -11,7 +11,8 @@
* use an empty (initialized) DB — they prove the lock is taken regardless of
* whether any rows match.
*/
import { describe, it, expect, vi, beforeEach } from 'vitest';
import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
import lbug from '@ladybugdb/core';
import { withTestLbugDB } from '../helpers/test-indexed-db.js';
import { NODE_TABLES } from '../../src/core/lbug/schema.js';
@ -40,6 +41,114 @@ import { closeQueryResults } from '../../src/core/lbug/query-result-utils.js';
const closeSpy = vi.mocked(closeQueryResults);
withTestLbugDB('conn-serialization', () => {
describe('incremental identity cursor lifetime', () => {
afterEach(() => vi.restoreAllMocks());
it('reads real rows without getAll and releases the cursor after a visitor failure', async () => {
const adapter = await import('../../src/core/lbug/lbug-adapter.js');
const getAll = vi.spyOn(lbug.QueryResult.prototype, 'getAll');
const close = vi.spyOn(lbug.QueryResult.prototype, 'close');
const seen: number[] = [];
await adapter.forEachQueryRow('UNWIND [1, 2, 3] AS n RETURN n', (row) => {
seen.push(row.n);
});
expect(seen).toEqual([1, 2, 3]);
expect(getAll).not.toHaveBeenCalled();
expect(close).toHaveBeenCalledTimes(1);
const error = new Error('bad identity');
await expect(
adapter.forEachQueryRow('RETURN 1 AS n', () => {
throw error;
}),
).rejects.toBe(error);
expect(close).toHaveBeenCalledTimes(2);
await expect(adapter.executeQuery('RETURN 2 AS n')).resolves.toEqual([{ n: 2 }]);
});
it('yields while retaining the lock through the last row and cursor close', async () => {
const adapter = await import('../../src/core/lbug/lbug-adapter.js');
let read = 0;
let visited = 0;
let yieldedAt = 0;
const close = vi.fn();
const cursor = {
hasNext: () => read < 2050,
getNextSync: () => {
expect(visited).toBe(read); // The previous row was consumed before reading another.
return { n: ++read };
},
close,
};
const query = vi.spyOn(lbug.Connection.prototype, 'query').mockImplementation(async (sql) => {
expect(sql).toBe('scan');
return cursor as any;
});
let checkpoint: Promise<unknown> | undefined;
// Schedule outside the reader's lock context, like the WAL driver timer.
setImmediate(() => {
yieldedAt = visited;
checkpoint = adapter.executeQuery('CHECKPOINT');
});
const prepare = vi.spyOn(lbug.Connection.prototype, 'prepare');
// executeQuery reaches prepare only once the reader has closed.
prepare.mockImplementation(async () => {
expect(visited).toBe(2050);
expect(close).toHaveBeenCalledTimes(1);
return { isSuccess: () => true } as any;
});
vi.spyOn(lbug.Connection.prototype, 'execute').mockResolvedValue({
getAll: async () => [],
close: () => {},
} as any);
await adapter.forEachQueryRow('scan', () => {
visited++;
});
await checkpoint;
expect(yieldedAt).toBeGreaterThan(0);
expect(yieldedAt).toBeLessThan(2050);
expect(query).toHaveBeenCalledTimes(1);
expect(prepare).toHaveBeenCalledTimes(1);
});
it.each(['none', 'multiple'] as const)(
'rejects %s result cursors and closes all returned cursors',
async (kind) => {
const { forEachQueryRow } = await import('../../src/core/lbug/lbug-adapter.js');
const close = vi.fn();
const cursors = kind === 'none' ? [] : [{ close }, { close }];
vi.spyOn(lbug.Connection.prototype, 'query').mockResolvedValue(cursors as any);
const visit = vi.fn();
await expect(forEachQueryRow('scan', visit)).rejects.toThrow('exactly one query result');
expect(visit).not.toHaveBeenCalled();
expect(close).toHaveBeenCalledTimes(cursors.length);
},
);
it.each(['read', 'null', 'close', 'read-and-close'] as const)(
'fails closed on %s failure',
async (stage) => {
const { forEachQueryRow } = await import('../../src/core/lbug/lbug-adapter.js');
const readError = new Error('native read failure');
const closeError = new Error('native close failure');
const close = vi.fn(() => {
if (stage.includes('close')) throw closeError;
});
vi.spyOn(lbug.Connection.prototype, 'query').mockResolvedValue({
hasNext: () => stage !== 'close',
getNextSync: () => {
if (stage === 'null') return null;
throw readError;
},
close,
} as any);
const outcome = forEachQueryRow('scan', vi.fn());
if (stage === 'null') await expect(outcome).rejects.toThrow('advertised next row');
else await expect(outcome).rejects.toBe(stage === 'close' ? closeError : readError);
expect(close).toHaveBeenCalledTimes(1);
},
);
});
describe('singleton-conn helpers acquire withConnLock (P1 #2264)', () => {
beforeEach(() => {
// Setup (clear/seed/flush) already exercised the lock; reset so each

View file

@ -52,7 +52,12 @@ it.each([
nativePhases,
size: { nodes: 1 },
reconcileGraphNodeIdentities,
adapter: { executeQuery: async (query: string) => (query.includes('Function') ? rows : []) },
adapter: {
executeQuery: async (query: string) => (query.includes('Function') ? rows : []),
forEachQueryRow: async (query: string, onRow: (row: (typeof rows)[number]) => void) => {
if (query.includes('Function')) rows.forEach(onRow);
},
},
}) as (phase: string) => Promise<void>;
await audit('controlled-scan');
expect(nativePhases).toEqual([expect.objectContaining({ verdict, rows: rows.length })]);

View file

@ -138,7 +138,7 @@ describe('incremental graph identity before publication', () => {
);
await adapter.tryFlushWAL();
await expect(
reconcileGraphNodeIdentities(graph, adapter.executeQuery, 'native CSV parity'),
reconcileGraphNodeIdentities(graph, adapter.forEachQueryRow, 'native CSV parity'),
).resolves.toMatchObject({ nodes: 5 });
} finally {
await adapter.closeLbug();
@ -264,7 +264,7 @@ describe('incremental graph identity before publication', () => {
await expect(
reconcileGraphNodeIdentities(
recovered.pipelineResult.graph,
adapter.executeQuery,
adapter.forEachQueryRow,
'recovered/reopened',
),
).resolves.toMatchObject({ nodes: expect.any(Number) });

View file

@ -0,0 +1,177 @@
import { execFileSync, spawnSync } from 'node:child_process';
import fs from 'node:fs';
import os from 'node:os';
import path from 'node:path';
import { fileURLToPath } from 'node:url';
import { afterEach, describe, expect, it } from 'vitest';
const harness = fileURLToPath(new URL('../../bench/lexical-lookup/measure.mjs', import.meta.url));
const scratch: string[] = [];
afterEach(() => {
for (const root of scratch.splice(0)) fs.rmSync(root, { recursive: true, force: true });
});
describe('lexical benchmark preflight', () => {
it.each([
['GITNEXUS_LEXICAL_BENCH_SCREEN', 'Screen must be 0 or 1'],
['GITNEXUS_LEXICAL_BENCH_BASE_FIRST', 'Base first must be 0 or 1'],
])('rejects invalid %s before touching corpora', (setting, message) => {
const root = fs.mkdtempSync(path.join(os.tmpdir(), 'lexical-option-'));
scratch.push(root);
const output = path.join(root, 'output');
const result = spawnSync(process.execPath, [harness, root, root, root, root, output], {
encoding: 'utf8',
env: { ...process.env, [setting]: 'false' },
});
expect(result.status).toBe(1);
expect(result.stderr).toContain(message);
expect(fs.existsSync(output)).toBe(false);
});
it.each(['empty', 'indexes', 'results.json'])(
'rejects existing %s output before touching corpora',
(contents) => {
const root = fs.mkdtempSync(path.join(os.tmpdir(), 'lexical-preflight-'));
scratch.push(root);
const output = path.join(root, 'output');
fs.mkdirSync(output);
if (contents === 'indexes') fs.mkdirSync(path.join(output, 'indexes'));
if (contents === 'results.json')
fs.writeFileSync(path.join(output, 'results.json'), 'prior result');
const before = path.join(root, 'before');
const after = path.join(root, 'after');
const result = spawnSync(process.execPath, [harness, root, root, before, after, output], {
encoding: 'utf8',
env: {
...process.env,
GITNEXUS_LEXICAL_BENCH_WORKER: '0',
GITNEXUS_LEXICAL_BENCH_WORKERS: '2',
GITNEXUS_LEXICAL_BENCH_HEAP_MIB: '4096',
},
});
expect(result.status).toBe(1);
expect(result.stderr).toContain('Output already exists');
expect(fs.existsSync(before)).toBe(false);
expect(fs.existsSync(after)).toBe(false);
expect(fs.readdirSync(output)).toEqual(contents === 'empty' ? [] : [contents]);
if (contents === 'results.json')
expect(fs.readFileSync(path.join(output, contents), 'utf8')).toBe('prior result');
},
);
});
describe('lexical benchmark sample collection', () => {
it.each(['fresh', 'persistent'])(
'collects a complete reversed-order baseline with a slower candidate (%s)',
(lifecycle) => {
const root = fs.mkdtempSync(path.join(os.tmpdir(), 'lexical-samples-'));
scratch.push(root);
const corpora = ['before', 'after'].map((name) => {
const repo = path.join(root, name);
fs.mkdirSync(repo);
execFileSync('git', ['init', '--quiet'], { cwd: repo });
execFileSync(
'git',
[
'-c',
'user.name=Benchmark Test',
'-c',
'user.email=test@example.test',
'-c',
'commit.gpgsign=false',
'commit',
'--quiet',
'--allow-empty',
'-m',
'fixture',
],
{
cwd: repo,
env: {
...process.env,
GIT_AUTHOR_DATE: '2020-01-01T00:00:00Z',
GIT_COMMITTER_DATE: '2020-01-01T00:00:00Z',
},
},
);
return repo;
});
const runtimes = [1, 100].map((duration) => {
const runtime = path.join(root, `runtime-${duration}`);
const core = path.join(runtime, 'gitnexus/dist/core');
fs.mkdirSync(path.join(core, 'lbug'), { recursive: true });
fs.writeFileSync(path.join(runtime, 'package.json'), '{"type":"module"}');
// A deterministic clock isolates collection/order behavior from machine load.
fs.writeFileSync(
path.join(core, 'run-analyze.js'),
`
import fs from 'node:fs';
import path from 'node:path';
let clock = 0;
Object.defineProperty(globalThis, 'performance', {value: {now: () => clock}});
export async function runFullAnalysis(repo) {
clock += ${duration};
const target = fs.readFileSync(path.join(repo, 'gitnexus/src/semantic-bench-control.ts'), 'utf8').includes('s.load()') ? 'load' : 'save';
return {
storagePath: repo, stats: {}, rebuildReasons: [],
incrementalStats: {writeMode: 'incremental', changedFiles: 1},
pipelineResult: {graph: {
*iterNodes() { yield {id: 'semanticBenchUse', label: 'Function'}; },
*iterRelationships() { yield {sourceId: 'semanticBenchUse', targetId: target, type: 'CALLS'}; }
}}
};
}
`,
);
fs.writeFileSync(path.join(core, 'logger.js'), 'export function configureLoggerSink() {}');
fs.writeFileSync(
path.join(core, 'lbug/lbug-adapter.js'),
`
import fs from 'node:fs';
import path from 'node:path';
let repo;
export async function initLbug(db) { repo = path.dirname(db); }
export async function executeQuery() {
return [{target: fs.readFileSync(path.join(repo, 'gitnexus/src/semantic-bench-control.ts'), 'utf8').includes('s.load()') ? 'load' : 'save'}];
}
export async function closeLbug() {}
`,
);
return runtime;
});
for (const screen of ['0', '1']) {
const output = path.join(root, `output-${screen}`);
const result = spawnSync(
process.execPath,
[harness, ...runtimes, ...corpora, output, lifecycle],
{
encoding: 'utf8',
env: {
...process.env,
GITNEXUS_LEXICAL_BENCH_WORKER: '0',
GITNEXUS_LEXICAL_BENCH_WORKERS: '2',
GITNEXUS_LEXICAL_BENCH_HEAP_MIB: '256',
GITNEXUS_LEXICAL_BENCH_SCREEN: screen,
GITNEXUS_LEXICAL_BENCH_BASE_FIRST: '0',
},
},
);
expect(result.status, result.stderr).toBe(0);
const report = JSON.parse(fs.readFileSync(path.join(output, 'results.json'), 'utf8'));
expect(report.screen).toBe(screen === '1');
expect(report.baseFirst).toBe(false);
expect(report.results.map((sample: { name: string }) => sample.name)).toEqual([
'candidate-seed',
'baseline-seed',
'candidate-1',
'baseline-1',
...(screen === '0' ? ['baseline-2', 'candidate-2', 'candidate-3', 'baseline-3'] : []),
]);
for (const repo of corpora)
expect(fs.existsSync(path.join(repo, 'gitnexus/src/semantic-bench-control.ts'))).toBe(
false,
);
}
},
);
});

View file

@ -120,13 +120,14 @@ const mockLbugAdapter = async () => {
initLbug: vi.fn(async () => undefined),
loadGraphToLbug: vi.fn(async () => undefined),
getLbugStats: vi.fn(async () => ({ nodes: 1, edges: 0, communities: 0, processes: 0 })),
executeQuery: vi.fn(async () => []),
// The policy fixture has one stored File row. Publication now reconciles
// that identity instead of trusting the synthetic node count alone.
executeQuery: vi.fn(async (query: string) =>
query.startsWith('MATCH (n:`File`) RETURN n.id AS id')
? [{ id: 'file:src/a.ts', name: REL_FILE, filePath: REL_FILE }]
: [],
),
forEachQueryRow: vi.fn<typeof actual.forEachQueryRow>(async (query, onRow) => {
if (query.startsWith('MATCH (n:`File`) RETURN n.id AS id')) {
onRow({ id: 'file:src/a.ts', name: REL_FILE, filePath: REL_FILE });
}
}),
executeWithReusedStatement: vi.fn(async () => []),
closeLbug: vi.fn(async () => undefined),
wipeLbugDbFiles: vi.fn(async () => undefined),

View file

@ -14,6 +14,7 @@ import {
findClassBindingInScope,
findReceiverTypeBinding,
isNamespaceNameShadowed,
lookupNameClaim,
} from '../../../src/core/ingestion/scope-resolution/scope/walkers.js';
import { findCallableBindingsAndAdlBlocker } from '../../../src/core/ingestion/languages/cpp/callable-bindings.js';
import { followChainPostFinalize } from '../../../src/core/ingestion/scope-resolution/passes/imported-return-types.js';
@ -60,6 +61,40 @@ function indexes(
}
describe('lexical name ownership', () => {
it.each(['Function', 'Class'] as const)(
'keeps %s lookup snapshots after source bindings change',
(kind) => {
const bindings = new Map<string, BindingRef[]>();
const inner = scope('fn', 'module', { bindings, kind });
const outer = scope('module', null);
const state = { ...indexes(inner, outer), bindings: new Map([[inner.id, bindings]]) };
const missing = lookupNameClaim(inner.id, 'helper', state);
expect(missing.status).toBe('absent');
const local: BindingRef[] = [binding];
bindings.set('helper', local);
const found = lookupNameClaim(inner.id, 'helper', state);
expect(found.bindings).toEqual([binding]);
local.push({ def: { ...target, nodeId: 'second:helper' }, origin: 'local' });
expect(found.bindings).toEqual([binding]);
expect(missing.bindings).toEqual([]);
expect(lookupNameClaim(inner.id, 'helper', state).bindings).toHaveLength(2);
},
);
it('clears deferred activation without mutating reused caller options', () => {
const late: BindingRef = { ...binding, availableFrom: { startLine: 8, startCol: 0 } };
const outer = scope('module', null, { bindings: new Map([['helper', [late]]]) });
const inner = scope('fn', outer.id, { lookupPolicy: { deferParentActivation: true } });
const options = Object.freeze({ position: Object.freeze({ startLine: 1, startCol: 0 }) });
const state = indexes(inner, outer);
expect(lookupNameClaim(inner.id, 'helper', state, options).bindings).toEqual([late]);
expect(lookupNameClaim(outer.id, 'helper', state, options).status).toBe('absent');
expect(
lookupNameClaim(inner.id, 'helper', state, { ...options, preservePosition: true }).status,
).toBe('absent');
expect(options.position).toEqual({ startLine: 1, startCol: 0 });
});
it('keeps touching half-open import ranges separate', () => {
const earlier = { startLine: 1, startCol: 0, endLine: 1, endCol: 10 };
const later = { startLine: 1, startCol: 10, endLine: 1, endCol: 20 };

View file

@ -41,7 +41,9 @@ const rows = () => [
},
];
const queryFor = (functions: ReturnType<typeof rows>) =>
vi.fn(async (query: string) => (query.includes('(n:`Function`)') ? functions : []));
vi.fn(async (query: string, onRow: (row: ReturnType<typeof rows>[number]) => void) => {
if (query.includes('(n:`Function`)')) functions.forEach(onRow);
});
describe('incremental identity reconciliation', () => {
it('certifies every identity using one unfiltered query per label', async () => {