@bldrs-ai/conway 1.559.1526 → 1.560.1539

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -30,7 +30,7 @@ var import_node_process = require("node:process");
30
30
  var readline = __toESM(require("node:readline"), 1);
31
31
 
32
32
  // compiled/src/version/version.js
33
- var versionString = "Conway v1.559.1526";
33
+ var versionString = "Conway v1.560.1539";
34
34
 
35
35
  // compiled/dependencies/conway-geom/interface/conway_geometry.js
36
36
  var wasmType = "";
@@ -14965,7 +14965,7 @@ ${t5.join("\n")}` : "";
14965
14965
  var import_process = require("process");
14966
14966
 
14967
14967
  // compiled/src/version/version.js
14968
- var versionString = "Conway v1.559.1526";
14968
+ var versionString = "Conway v1.560.1539";
14969
14969
 
14970
14970
  // compiled/dependencies/conway-geom/interface/conway_geometry.js
14971
14971
  function pThreadsAllowed() {
@@ -15943,7 +15943,7 @@ var ParsingBuffer = class {
15943
15943
  };
15944
15944
 
15945
15945
  // compiled/src/version/version.js
15946
- var versionString = "Conway v1.559.1526";
15946
+ var versionString = "Conway v1.560.1539";
15947
15947
 
15948
15948
  // compiled/dependencies/conway-geom/interface/conway_geometry.js
15949
15949
  function pThreadsAllowed() {
@@ -944,7 +944,7 @@ var EntityTypesIfcCount = 909;
944
944
  var entity_types_ifc_gen_default = EntityTypesIfc;
945
945
 
946
946
  // compiled/src/version/version.js
947
- var versionString = "Conway v1.559.1526";
947
+ var versionString = "Conway v1.560.1539";
948
948
 
949
949
  // compiled/dependencies/conway-geom/interface/conway_geometry.js
950
950
  var wasmType = "";
@@ -87,8 +87,16 @@ export declare function canSettleMemory(): boolean;
87
87
  * produces `parseTimeMs` / `geometryTimeMs` / `totalTimeMs` — baseline before
88
88
  * the load starts, retained sample after teardown. That property is what
89
89
  * makes the retention columns free, and it is checkable by measurement: a run
90
- * with `--expose-gc` and a run without share identical code, so if the timing
91
- * columns move between them, the settle is leaking into the measured window.
90
+ * with `--expose-gc` and a run without share identical code, so any movement
91
+ * in the timing columns between them is down to the settle. The SIGN says
92
+ * which way: gc-on slower is the settle leaking into the measured window,
93
+ * which is a bug; gc-on faster is the pre-load settle taking engine-init
94
+ * garbage OUT of that window. That is an absolute cost of about 10 ms per
95
+ * load, measured as 13-16% of `parseTimeMs` on models that parse in ~60 ms
96
+ * and unresolvable against one that parses in 578 ms, so read a ratio
97
+ * against the model's own parse time. `rc-regression.yml` runs both
98
+ * conditions in one job; see design/new/perf-measurement.md §"The settle
99
+ * also cleans the window".
92
100
  *
93
101
  * @return {Promise<SettledMemorySample | undefined>} The settled sample, or
94
102
  * undefined where no collector is exposed — in which case the caller emits
@@ -1 +1 @@
1
- {"version":3,"file":"retained_memory.d.ts","sourceRoot":"","sources":["../../../src/core/retained_memory.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AASH;;;;;;;GAOG;AACH,MAAM,WAAW,mBAAmB;IAElC,kCAAkC;IAClC,QAAQ,EAAE,MAAM,CAAA;IAEhB,8CAA8C;IAC9C,aAAa,EAAE,MAAM,CAAA;IAErB,uDAAuD;IACvD,aAAa,EAAE,MAAM,CAAA;CACtB;AAED,8DAA8D;AAC9D,MAAM,WAAW,gBAAgB;IAE/B,2DAA2D;IAC3D,KAAK,EAAE,MAAM,CAAA;IAEb,oEAAoE;IACpE,UAAU,EAAE,MAAM,CAAA;IAElB,sEAAsE;IACtE,UAAU,EAAE,MAAM,CAAA;CACnB;AAED;;;;;;;;GAQG;AACH,wBAAgB,SAAS,IAAI,CAAE,MAAM,IAAI,CAAE,GAAG,SAAS,CAKtD;AAED;;;;;;;;;;;;GAYG;AACH,wBAAgB,eAAe,IAAI,OAAO,CAKzC;AAED;;;;;;;;;;;;;;;;;;;;;;GAsBG;AACH,wBAAsB,qBAAqB,IACvC,OAAO,CAAC,mBAAmB,GAAG,SAAS,CAAC,CAuB3C;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAsB,4BAA4B,CAC9C,QAAQ,EAAE,MAAM,GAAI,OAAO,CAAC,mBAAmB,GAAG,SAAS,CAAC,CAG/D;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,gBAAgB,CAC5B,QAAQ,EAAE,mBAAmB,GAAG,SAAS,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,GAAI,gBAAgB,GAAG,SAAS,CAa5E"}
1
+ {"version":3,"file":"retained_memory.d.ts","sourceRoot":"","sources":["../../../src/core/retained_memory.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AASH;;;;;;;GAOG;AACH,MAAM,WAAW,mBAAmB;IAElC,kCAAkC;IAClC,QAAQ,EAAE,MAAM,CAAA;IAEhB,8CAA8C;IAC9C,aAAa,EAAE,MAAM,CAAA;IAErB,uDAAuD;IACvD,aAAa,EAAE,MAAM,CAAA;CACtB;AAED,8DAA8D;AAC9D,MAAM,WAAW,gBAAgB;IAE/B,2DAA2D;IAC3D,KAAK,EAAE,MAAM,CAAA;IAEb,oEAAoE;IACpE,UAAU,EAAE,MAAM,CAAA;IAElB,sEAAsE;IACtE,UAAU,EAAE,MAAM,CAAA;CACnB;AAED;;;;;;;;GAQG;AACH,wBAAgB,SAAS,IAAI,CAAE,MAAM,IAAI,CAAE,GAAG,SAAS,CAKtD;AAED;;;;;;;;;;;;GAYG;AACH,wBAAgB,eAAe,IAAI,OAAO,CAKzC;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AACH,wBAAsB,qBAAqB,IACvC,OAAO,CAAC,mBAAmB,GAAG,SAAS,CAAC,CAuB3C;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAsB,4BAA4B,CAC9C,QAAQ,EAAE,MAAM,GAAI,OAAO,CAAC,mBAAmB,GAAG,SAAS,CAAC,CAG/D;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,gBAAgB,CAC5B,QAAQ,EAAE,mBAAmB,GAAG,SAAS,EACzC,QAAQ,EAAE,mBAAmB,GAAG,SAAS,GAAI,gBAAgB,GAAG,SAAS,CAa5E"}
@@ -72,8 +72,16 @@ export function canSettleMemory() {
72
72
  * produces `parseTimeMs` / `geometryTimeMs` / `totalTimeMs` — baseline before
73
73
  * the load starts, retained sample after teardown. That property is what
74
74
  * makes the retention columns free, and it is checkable by measurement: a run
75
- * with `--expose-gc` and a run without share identical code, so if the timing
76
- * columns move between them, the settle is leaking into the measured window.
75
+ * with `--expose-gc` and a run without share identical code, so any movement
76
+ * in the timing columns between them is down to the settle. The SIGN says
77
+ * which way: gc-on slower is the settle leaking into the measured window,
78
+ * which is a bug; gc-on faster is the pre-load settle taking engine-init
79
+ * garbage OUT of that window. That is an absolute cost of about 10 ms per
80
+ * load, measured as 13-16% of `parseTimeMs` on models that parse in ~60 ms
81
+ * and unresolvable against one that parses in 578 ms, so read a ratio
82
+ * against the model's own parse time. `rc-regression.yml` runs both
83
+ * conditions in one job; see design/new/perf-measurement.md §"The settle
84
+ * also cleans the window".
77
85
  *
78
86
  * @return {Promise<SettledMemorySample | undefined>} The settled sample, or
79
87
  * undefined where no collector is exposed — in which case the caller emits
@@ -29,9 +29,18 @@ const CHILD_NODE_FLAGS = '--experimental-specifier-resolution=node';
29
29
  * both samples sit outside the timed region — baseline before the load,
30
30
  * retained sample after teardown — so a run with the flag and a run without
31
31
  * should produce the same timing columns from identical code. Set the
32
- * variable to `0`, `false` or `off` for the without side of that A/B. If the
33
- * timing columns move between the two, the settle is reaching the measured
34
- * window, and that is a bug rather than a tolerance.
32
+ * variable to `0`, `false` or `off` for the without side of that A/B; the
33
+ * `rebless` job of `.github/workflows/rc-regression.yml` does exactly that
34
+ * for its second, control pass, because the two conditions have to share a
35
+ * runner for the comparison to mean anything (between two CI runs the timing
36
+ * columns carry a ~1.5x runner scale factor against a ~1% effect).
37
+ *
38
+ * Read any movement with its sign. Gc-on SLOWER means the settle is reaching
39
+ * the measured window, which is a bug rather than a tolerance. Gc-on FASTER
40
+ * means the opposite — the pre-load settle collects engine-init garbage that
41
+ * the flag-off run collects inside the timed region instead. That is about
42
+ * 10 ms per load: 13-16% of `parseTimeMs` on a model that parses in ~60 ms,
43
+ * and lost in the noise on one that parses in 578 ms.
35
44
  *
36
45
  * @return {boolean} True unless CONWAY_PERF_EXPOSE_GC disables it.
37
46
  */
@@ -16,7 +16,7 @@ import { createRequire } from 'module';
16
16
  const require_ = createRequire(import.meta.url);
17
17
  // Resolved from the repo root: the test runs from compiled/src/scripts, and
18
18
  // scripts/ is not part of the tsc build. Jest's rootDir is the repo root.
19
- const { DETAIL_COLUMNS, findPreviousSnapshot, isChronologicalDelta, removeStaleDeltas, writeDetailCsv, versionCompare, } = require_(path.resolve(process.cwd(), 'scripts/bless_perf_snapshot.cjs'));
19
+ const { DETAIL_COLUMNS, findPreviousSnapshot, isChronologicalDelta, removeStaleDeltas, renderReadme, writeDetailCsv, versionCompare, } = require_(path.resolve(process.cwd(), 'scripts/bless_perf_snapshot.cjs'));
20
20
  const { parseCsv } = require_(path.resolve(process.cwd(), 'scripts/csv_rfc4180.cjs'));
21
21
  let workDir;
22
22
  beforeEach(() => {
@@ -328,3 +328,90 @@ describe('isChronologicalDelta', () => {
328
328
  }
329
329
  });
330
330
  });
331
+ describe('renderReadme', () => {
332
+ const info = {
333
+ engine: 'conway1.560.1600-ci',
334
+ repoName: 'test-models',
335
+ modelCount: 97,
336
+ retentionCount: 97,
337
+ deltaName: 'conway1.451.1357-ci_1.560.1600_delta.csv',
338
+ previousName: 'conway1.451.1357-ci_test-models',
339
+ };
340
+ test('does not repeat the pre-#553 claim that geometryMemoryMb is N/A', () => {
341
+ // #553 restored geometryMemoryMb on the conway-native writer (SKYLARK250
342
+ // reads 185.22 against the 185.836 recorded at 1.451). The README shipped
343
+ // beside the 1.549 snapshot still says it is unmeasured, which is now the
344
+ // opposite of the truth for a column carrying real data; that text was
345
+ // removed from this template, and this keeps it out.
346
+ const text = renderReadme(info);
347
+ expect(text).not.toContain('`geometryMemoryMb` is absent here');
348
+ expect(text).toContain('`schemaVersion`,\n`preprocessorVersion` and `originatingSystem` are `N/A`');
349
+ // The N/A-is-not-a-zero framing is the point of #548 and must survive for
350
+ // the columns genuinely absent from an older snapshot.
351
+ expect(text).toContain('That is a missing\nmeasurement, not a zero');
352
+ });
353
+ test('says on its own face whether the settle ran', () => {
354
+ // A directory of N/A retention must explain itself without anyone having
355
+ // to find the workflow that produced it.
356
+ expect(renderReadme(info)).toContain('Retention is measured on 97 of 97');
357
+ expect(renderReadme({ ...info, retentionCount: 0 }))
358
+ .toContain('Retention is `N/A` on every row here');
359
+ });
360
+ test('names the blessed pass as the one this snapshot came from', () => {
361
+ const text = renderReadme(info);
362
+ expect(text).toContain('CONWAY_PERF_EXPOSE_GC=0');
363
+ expect(text).toContain('The control pass is never blessed');
364
+ });
365
+ test('the N/A inventory covers the retention columns, not just FAIL rows', () => {
366
+ // A settle-less snapshot used to assert both 'Every other column is
367
+ // measured - with the exception of a row whose loadStatus is not OK' and
368
+ // 'Retention is `N/A` on every row here', four paragraphs apart in one
369
+ // file. The retention columns read N/A on an OK row whenever the run had
370
+ // no --expose-gc, which is exactly the run this branch describes.
371
+ const text = renderReadme({ ...info, retentionCount: 0 });
372
+ expect(text).toContain('Retention is `N/A` on every row here');
373
+ expect(text).not.toContain('Every other column is measured — with the');
374
+ expect(text).toContain('the three retention columns carry `N/A`');
375
+ expect(text).toContain('on an `OK` row as\nmuch as a failed one');
376
+ });
377
+ test('attributes the geometryMemoryMb split to the writers #555 measured', () => {
378
+ // #555 measured 16.8 vs 22.3 MB between `ifc_command_line_main` and the
379
+ // IFC regression child - two IFC pipelines. MB-Khaya never reaches the
380
+ // AP214 child, so pinning that figure to the IFC-row-vs-STEP-row split
381
+ // this file mixes would cite evidence for a claim it does not support.
382
+ const text = renderReadme(info);
383
+ expect(text).toContain('The IFC **CLI** and the IFC\nregression child read 16.8 vs 22.3 MB');
384
+ expect(text).toContain('has not been measured to disagree');
385
+ });
386
+ test('describes the second-engine term as closed by #557, not as current', () => {
387
+ // conway#557 landed: both regression children now extract on the engine
388
+ // main() initialised. A README still saying an IFC row carries ~100 MB of
389
+ // second engine would be describing a world that no longer exists - and
390
+ // the boundary that DOES matter now is that pre-#557 snapshots carry the
391
+ // constant and this one does not.
392
+ const text = renderReadme(info);
393
+ expect(text).toContain('carried a second, unrelated split until conway#557');
394
+ expect(text).toContain('~55-60 MB constant on every IFC row');
395
+ expect(text).not.toContain('roughly 100 MB');
396
+ expect(text).toContain('A snapshot blessed\nbefore conway#557');
397
+ });
398
+ test('warns off the misreadings the columns invite', () => {
399
+ const text = renderReadme(info);
400
+ // Retention is live model + leak, not leak.
401
+ expect(text).toContain('still-live model plus anything genuinely leaked');
402
+ // The two columns that are pipeline-scoped, with their issues.
403
+ expect(text).toContain('conway/issues/555');
404
+ expect(text).toContain('conway/issues/557');
405
+ // The third native quantity, and the arrayBuffers/external subset rule.
406
+ expect(text).toContain('getAllocationSize');
407
+ expect(text).toContain('ArrayBuffer subset');
408
+ // Cross-run timing carries the runner scale factor.
409
+ expect(text).toContain('median 1.55x');
410
+ });
411
+ test('names the delta and its predecessor, or says there is none', () => {
412
+ expect(renderReadme(info)).toContain('`conway1.451.1357-ci_1.560.1600_delta.csv` diffs this run against ' +
413
+ '`../conway1.451.1357-ci_test-models/`.');
414
+ expect(renderReadme({ ...info, deltaName: '', previousName: '' }))
415
+ .toContain('no delta was produced');
416
+ });
417
+ });
@@ -0,0 +1,2 @@
1
+ export {};
2
+ //# sourceMappingURL=perf_ab_compare.test.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"perf_ab_compare.test.d.ts","sourceRoot":"","sources":["../../../src/scripts/perf_ab_compare.test.ts"],"names":[],"mappings":""}
@@ -0,0 +1,218 @@
1
+ import fs from 'fs';
2
+ import os from 'os';
3
+ import path from 'path';
4
+ import { afterEach, beforeEach, describe, expect, test } from '@jest/globals';
5
+ import { createRequire } from 'module';
6
+ /**
7
+ * The rc-regression GC-settle A/B comparator (scripts/perf_ab_compare.cjs).
8
+ *
9
+ * The rc job runs the corpus twice in one job — the blessed pass in the
10
+ * shipped configuration, then a control pass with `CONWAY_PERF_EXPOSE_GC=0` —
11
+ * and differences them there, because between two separate CI runs the timing
12
+ * columns carry a ~1.5x runner scale factor two orders of magnitude larger
13
+ * than the ~1% effect under test (conway#554). This pins the three things
14
+ * that make the resulting number readable: that a control pass which measured
15
+ * retention is called out as an invalid A/B rather than reported as a result,
16
+ * that `N/A` propagates instead of being coerced to 0 (#548), and that
17
+ * `Date.now()`-quantised rows are excluded from the ratio statistics rather
18
+ * than allowed to dominate them.
19
+ */
20
+ /* eslint-disable no-magic-numbers -- the literals here are the fixture
21
+ timings and percentile fractions the cases are about; naming a 0.1 or a
22
+ 110 ms parse would hide the shape each case exists to express. */
23
+ const require_ = createRequire(import.meta.url);
24
+ const { RATIO_FLOOR_MS, checkSwitch, compareRuns, numeric, percentile, readPerfCsv, renderCsv, renderMarkdown, summariseColumn, } = require_(path.resolve(process.cwd(), 'scripts/perf_ab_compare.cjs'));
25
+ /**
26
+ * A perf.csv row with the columns this comparator reads, overridable per test.
27
+ *
28
+ * @param overrides Column values to set.
29
+ * @return {PerfRow} The row.
30
+ */
31
+ function row(overrides) {
32
+ return {
33
+ status: 'OK',
34
+ parseTimeMs: '100',
35
+ geometryTimeMs: '1000',
36
+ totalTimeMs: '1100',
37
+ heapUsedMb: '50.00',
38
+ rssMb: '300.00',
39
+ peakRssMb: '310.00',
40
+ retainedRssMb: 'N/A',
41
+ retainedHeapUsedMb: 'N/A',
42
+ retainedExternalMb: 'N/A',
43
+ ...overrides,
44
+ };
45
+ }
46
+ /**
47
+ * A blessed-pass row: the settle ran, so retention is measured.
48
+ *
49
+ * @param overrides Column values to set.
50
+ * @return {PerfRow} The row.
51
+ */
52
+ function blessedRow(overrides) {
53
+ return row({
54
+ retainedRssMb: '380.00',
55
+ retainedHeapUsedMb: '10.00',
56
+ retainedExternalMb: '47.00',
57
+ ...overrides,
58
+ });
59
+ }
60
+ describe('readPerfCsv', () => {
61
+ let workDir;
62
+ beforeEach(() => {
63
+ workDir = fs.mkdtempSync(path.join(os.tmpdir(), 'perf-ab-'));
64
+ });
65
+ afterEach(() => {
66
+ fs.rmSync(workDir, { recursive: true, force: true });
67
+ });
68
+ test('reads the 16-column perf.csv both passes write', () => {
69
+ // The exact header ifc_regression_main.ts emits, with a control-pass row:
70
+ // the file this comparator is pointed at in CI, not a reduced stand-in.
71
+ const file = path.join(workDir, 'perf-nogc.csv');
72
+ fs.writeFileSync(file, 'file,status,parseTimeMs,geometryTimeMs,totalTimeMs,geometryMemoryMb,' +
73
+ 'peakWasmHeapMb,rssMb,peakRssMb,heapUsedMb,heapTotalMb,externalMb,' +
74
+ 'arrayBuffersMb,retainedRssMb,retainedHeapUsedMb,retainedExternalMb\n' +
75
+ 'AC20-FZK-Haus.ifc,OK,76,436,512,1.69,35.13,289.39,289.52,54.40,84.34,' +
76
+ '9.98,7.93,N/A,N/A,N/A\n');
77
+ const rows = readPerfCsv(file);
78
+ expect(rows).toHaveLength(1);
79
+ expect(rows[0].file).toBe('AC20-FZK-Haus.ifc');
80
+ expect(rows[0].totalTimeMs).toBe('512');
81
+ expect(rows[0].retainedRssMb).toBe('N/A');
82
+ });
83
+ test('an empty file is no rows, not a throw', () => {
84
+ const file = path.join(workDir, 'empty.csv');
85
+ fs.writeFileSync(file, '');
86
+ expect(readPerfCsv(file)).toEqual([]);
87
+ });
88
+ });
89
+ describe('numeric', () => {
90
+ test('treats N/A and blank as absent rather than zero', () => {
91
+ // #548: parseValue returning 0.0 for a missing column produced a phantom
92
+ // -185.836 MB "win" for SKYLARK250. Nothing in this file may repeat it.
93
+ expect(numeric('N/A')).toBeNull();
94
+ expect(numeric('')).toBeNull();
95
+ expect(numeric(undefined)).toBeNull();
96
+ expect(numeric('0')).toBe(0);
97
+ expect(numeric('12.5')).toBe(12.5);
98
+ });
99
+ });
100
+ describe('percentile', () => {
101
+ test('interpolates between samples and handles the empty case', () => {
102
+ expect(percentile([1, 2, 3, 4], 0.5)).toBeCloseTo(2.5);
103
+ expect(percentile([1, 2, 3, 4, 5], 0.1)).toBeCloseTo(1.4);
104
+ expect(percentile([], 0.5)).toBeNull();
105
+ });
106
+ });
107
+ describe('checkSwitch', () => {
108
+ test('a control pass with no retention against a blessed pass with it is valid', () => {
109
+ const check = checkSwitch([blessedRow({ file: 'a.ifc' }), blessedRow({ file: 'b.ifc' })], [row({ file: 'a.ifc' }), row({ file: 'b.ifc' })]);
110
+ expect(check.valid).toBe(true);
111
+ expect(check.blessedMeasured).toBe(2);
112
+ expect(check.controlMeasured).toBe(0);
113
+ });
114
+ test('a control pass that measured retention is not an A/B at all', () => {
115
+ // The env switch never reached the children, so both passes ran the same
116
+ // configuration and every timing ratio below is noise against noise.
117
+ const check = checkSwitch([blessedRow({ file: 'a.ifc' })], [blessedRow({ file: 'a.ifc' })]);
118
+ expect(check.valid).toBe(false);
119
+ expect(check.controlMeasured).toBe(1);
120
+ });
121
+ test('a blessed pass with no retention is equally not an A/B', () => {
122
+ const check = checkSwitch([row({ file: 'a.ifc' })], [row({ file: 'a.ifc' })]);
123
+ expect(check.valid).toBe(false);
124
+ });
125
+ });
126
+ describe('summariseColumn', () => {
127
+ const pairs = (blessed, control) => blessed.map((value, index) => ({
128
+ file: `m${index}.ifc`,
129
+ blessed: row({ file: `m${index}.ifc`, ...value }),
130
+ control: row({ file: `m${index}.ifc`, ...control[index] }),
131
+ }));
132
+ test('ratio, sign count and median delta', () => {
133
+ const stat = summariseColumn(pairs([{ totalTimeMs: '110' }, { totalTimeMs: '90' }, { totalTimeMs: '100' }], [{ totalTimeMs: '100' }, { totalTimeMs: '100' }, { totalTimeMs: '100' }]), 'totalTimeMs', true);
134
+ expect(stat.n).toBe(3);
135
+ expect(stat.medianRatio).toBeCloseTo(1.0);
136
+ expect(stat.slower).toBe(1);
137
+ expect(stat.medianDelta).toBe(0);
138
+ });
139
+ test('rows under the quantisation floor are counted out, not counted in', () => {
140
+ // A 3 ms parse carries a +/-1 ms Date.now() tick — a 33% "ratio" that is
141
+ // pure rounding, and there are enough tiny models in the corpus to swamp
142
+ // the median with them.
143
+ const stat = summariseColumn(pairs([{ parseTimeMs: '3' }, { parseTimeMs: '4' }, { parseTimeMs: '200' }], [{ parseTimeMs: '1' }, { parseTimeMs: '2' }, { parseTimeMs: '100' }]), 'parseTimeMs', true);
144
+ expect(RATIO_FLOOR_MS).toBe(10);
145
+ expect(stat.floored).toBe(2);
146
+ expect(stat.n).toBe(1);
147
+ expect(stat.medianRatio).toBeCloseTo(2.0);
148
+ });
149
+ test('a zero on the control side is floored, not dropped uncounted', () => {
150
+ // 0 ms is BELOW the 10 ms floor, so it belongs in `floored`. Testing
151
+ // `off === 0` ahead of the floor dropped such a row out of both counts,
152
+ // leaving n + floored unable to account for every pair - which the
153
+ // module header promises they can. Reachable: box.ifc parses in 1 ms in
154
+ // the committed snapshots, one Date.now() tick from 0, and a
155
+ // zero-geometry model's geometryTimeMs is the same story.
156
+ const stat = summariseColumn(pairs([{ geometryTimeMs: '0' }, { geometryTimeMs: '200' }], [{ geometryTimeMs: '0' }, { geometryTimeMs: '100' }]), 'geometryTimeMs', true);
157
+ expect(stat.floored).toBe(1);
158
+ expect(stat.n).toBe(1);
159
+ expect(stat.floored + stat.n).toBe(2);
160
+ });
161
+ test('the floor is not applied to the memory columns', () => {
162
+ const stat = summariseColumn(pairs([{ heapUsedMb: '5.0' }], [{ heapUsedMb: '8.0' }]), 'heapUsedMb', false);
163
+ expect(stat.floored).toBe(0);
164
+ expect(stat.n).toBe(1);
165
+ });
166
+ test('an unmeasured cell on either side drops the pair, it does not read zero', () => {
167
+ const stat = summariseColumn(pairs([{ totalTimeMs: 'N/A' }, { totalTimeMs: '200' }], [{ totalTimeMs: '100' }, { totalTimeMs: 'N/A' }]), 'totalTimeMs', true);
168
+ expect(stat.n).toBe(0);
169
+ expect(stat.medianRatio).toBeNull();
170
+ });
171
+ });
172
+ describe('compareRuns', () => {
173
+ test('joins on file, and reports rows only one pass produced', () => {
174
+ const comparison = compareRuns([blessedRow({ file: 'a.ifc' }), blessedRow({ file: 'only-blessed.ifc' })], [row({ file: 'a.ifc' }), row({ file: 'only-control.ifc' })]);
175
+ expect(comparison.pairs.map((pair) => pair.file)).toEqual(['a.ifc']);
176
+ expect(comparison.blessedOnly).toEqual(['only-blessed.ifc']);
177
+ expect(comparison.controlOnly).toEqual(['only-control.ifc']);
178
+ });
179
+ test('a model that failed in either pass is not timed against one that did not', () => {
180
+ const comparison = compareRuns([blessedRow({ file: 'a.ifc', status: 'FAIL' })], [row({ file: 'a.ifc' })]);
181
+ expect(comparison.pairs).toHaveLength(0);
182
+ });
183
+ test('summarises the three timing columns and the memory columns', () => {
184
+ const comparison = compareRuns([blessedRow({ file: 'a.ifc' })], [row({ file: 'a.ifc' })]);
185
+ expect(comparison.timing.map((stat) => stat.column))
186
+ .toEqual(['parseTimeMs', 'geometryTimeMs', 'totalTimeMs']);
187
+ expect(comparison.memory.map((stat) => stat.column))
188
+ .toEqual(['heapUsedMb', 'rssMb', 'peakRssMb']);
189
+ });
190
+ });
191
+ describe('renderMarkdown', () => {
192
+ test('a valid A/B says so, and reports the sign rule in both directions', () => {
193
+ const text = renderMarkdown(compareRuns([blessedRow({ file: 'a.ifc' })], [row({ file: 'a.ifc' })]), 'bldrs-ai/test-models');
194
+ expect(text).toContain('**Switch check: OK.**');
195
+ expect(text).toContain('gc on SLOWER');
196
+ expect(text).toContain('gc on FASTER');
197
+ });
198
+ test('an invalid A/B is labelled INVALID rather than presented as a result', () => {
199
+ const text = renderMarkdown(compareRuns([blessedRow({ file: 'a.ifc' })], [blessedRow({ file: 'a.ifc' })]), 'bldrs-ai/test-models');
200
+ expect(text).toContain('INVALID');
201
+ expect(text).toContain('answers nothing about the settle');
202
+ });
203
+ });
204
+ describe('renderCsv', () => {
205
+ test('carries both passes, the ratio, and the blessed retention columns', () => {
206
+ const text = renderCsv(compareRuns([blessedRow({ file: 'a.ifc', totalTimeMs: '110' })], [row({ file: 'a.ifc', totalTimeMs: '100' })]));
207
+ const [header, dataRow] = text.trim().split('\n');
208
+ expect(header).toContain('totalTimeMs_gcOn,totalTimeMs_gcOff,totalTimeMs_ratio');
209
+ expect(header).toContain('retainedRssMb_gcOn');
210
+ expect(dataRow).toContain('110,100,1.1000');
211
+ expect(dataRow).toContain('380.00');
212
+ });
213
+ test('an unmeasured input yields N/A, never a fabricated ratio', () => {
214
+ const text = renderCsv(compareRuns([blessedRow({ file: 'a.ifc', totalTimeMs: 'N/A' })], [row({ file: 'a.ifc', totalTimeMs: '100' })]));
215
+ const dataRow = text.trim().split('\n')[1];
216
+ expect(dataRow).toContain('N/A,100,N/A');
217
+ });
218
+ });
@@ -5,5 +5,5 @@
5
5
  // only the first segment (major) is meaningful and is the one CI carries forward.
6
6
  // Must stay in `vN.N.N` shape: the CI stamp regex, scripts/updateVersion.mjs, and
7
7
  // statistics.ts all match `v\d+\.\d+\.\d+`.
8
- const versionString = 'Conway v1.559.1526';
8
+ const versionString = 'Conway v1.560.1539';
9
9
  export { versionString };