gstack/docs/designs/gate-diff-calibration.md

gate-diff calibration — tool inventory over gstack's last 50 merged PRs

Generated by bun scripts/gate-diff-calibrate.ts --ref origin/main --count 50 --markdown on 2026-10-10 against origin/main at 20eb620 (v1.91.68.0), scanning each commit's first-parent delta exactly as gstack-gate-diff --commit <sha> does (regenerated after the RH-15? demotion and the shallow-boundary parent fix, so the table reflects the shipped detector). gstack squash-merges, so the 50 first-parent commits are the 50 landed PRs. The clone was shallow; git fetch --deepen=60 origin main was needed first (the tool reports a missing parent as GATE_ERROR=no_base … fix=git fetch --deepen).

What this measures, and what it does not

This is the tool's inventory: how many hunks the floor lists per PR, how many are eligible for the read level, how often the 50-hunk / 200 KB read cap makes a result partial, and how often each tag fires. Tags are factual labels (a test deleted beside its deleted source is still RH-1), so a tag count is not a count of problems.

It does not measure agent findings, kept/restored/open dispositions, or whether a reviewer would have wanted to read a tagged hunk. The design's 10-PR hand read of tagged hunks is a human step this table only supports; the agent-level signal is collected after ship from the persisted review records. No artifact was written: calibration never touches the state root.

Observations from the numbers

• Listed per PR: median 40.5, p75 267, max 1,276 (the v1.91.8.0 test-audit wave); eligible median 24.5. This matches the design's own measurement (median 23–45) and confirms per-hunk disposition is unworkable; the three levels are the right unit.
• Partial results: the read cap was hit in 20 of 50 PRs. For gstack's test-heavy PRs, "partial (K of M eligible read)" will be the common state, not the exception.
• RH-15 (paired, 35 hunks / 14 PRs) reads well by inspection: real timeout-minutes bumps in workflows, timeout 5000 -> 30000 in tests, a judge threshold 4 -> 3, an exact-count assertion relaxed to a bound, plus tightenings labelled as such. The direction note makes the read fast.
• RH-15? (unpaired): as first specified ("unpaired values are tagged RH-15? and read") the tag fired on 848 hunks across 41 PRs, dominated by spawnSync(…, { timeout: 30_000 }) in new or moved test code; this repository requires a timeout on every synchronous spawn (test/spawnsync-timeout-tripwire.test.ts), so nearly every test PR carried dozens of noise tags. Decision (integrator, per the design's demotion rule): RH-15? applies only to CI and runner/lint-config files; in test files unpaired values are inventory and only paired RH-15 is tagged. With that rule the table below shows 30 hunks across 13 PRs, all in workflow or runner config. The detector was also tightened to key-value assignments (key: N, key = N, key(N), --key N, and a gate key opening a nested structure) and to numeric literals with _ separators; prose strings and setTimeout calls no longer fire.
• RH-1 (411 hunks / 18 PRs): 343 of them are the v1.91.8.0 test-retirement PR; 29 are skip/only idioms (several are conditional process.platform === 'linux' ? describe : describe.skip). Factual, and exactly the inventory a reviewer of that PR would want listed.
• RH-16 (80 hunks / 19 PRs): gstack's many best-effort catch {} blocks. Overlaps slop-scan as the design states; listed under the RH id.
• RH-12, RH-13, RH-14: rare (6 / 8 / 3 hunks), plausible on inspection.
• RH-3 / RH-3?: did not fire; gstack has no __snapshots__/.golden fixtures in this window (its SKILL.md goldens are .md, listed by path only).
• Prose (.md, .md.tmpl) and generated/vendored paths (dist/, vendor/, lockfiles, .min.js) are floor inventory only; without that rule the design document itself tagged every RH id it describes.

Table

commit	subject	listed	eligible	read	unread	unmatched	unsupported	tags
20eb6202fa	v1.91.68.0 fix(cso-eval): Claude producers keep the source sealed, run t	18	16	16	0	6	0	-
54efba6dd5	v1.91.67.0 feat(hosts): Antigravity CLI (agy) host (#3089)	14	7	7	0	15	0	-
8bb6d65270	v1.91.66.0 fix: /autoplan without permission prompts after background re	133	88	50	38	119	0	RH-1:2 RH-15?:1 RH-16:2
b357b50378	v1.91.65.0 fix(install): skills only for the agents you installed, every	12	8	8	0	19	0	-
9e9e674b1e	v1.91.62.0 fix(cso): Rails verification fits two arm64 offline preparati	1	1	1	0	5	0	-
4dfd83bf8c	v1.91.60.0 fix(windows): setup degrades instead of dying on a CSO failur	25	8	8	0	17	0	RH-15:1 RH-16:2
4e5ffbb857	v1.91.59.0 fix(autoplan): bind init through symlinked plan directories a	6	4	4	0	8	0	-
9a1dc81a2b	v1.91.54.0 fix: office-hours first sessions get the first-session closin	13	6	6	0	18	0	-
d082eccc50	v1.91.52.0 fix(qa): annotations limits shape; qa-quick does less work in	2	1	1	0	7	0	-
c7edf46bcf	v1.91.47.0 fix(careful,freeze): safety hooks cover PowerShell and Notebo	11	7	7	0	22	0	-
92cfd07a79	v1.91.45.0 feat: configurable model tiers and advisory freshness (#3073)	52	26	26	0	60	0	RH-15?:1 RH-16:1
52eec18c63	v1.91.42.0 fix(cso): lift the snapshot file-count ceiling, bound report	2	1	1	0	9	0	RH-16:1
c7554824cb	v1.91.38.0 fix(ios-qa): iPads work, a dropped route no longer relaunches	17	10	10	0	13	0	-
1d1ce39b4e	v1.91.36.0 fix(cso): Python and Rails repairs replay, Rails verification	9	4	4	0	11	0	-
f67c478f05	v1.91.34.0 fix: /autoplan on current Claude Code and long sessions, quie	194	126	50	76	150	0	RH-1:1 RH-15:1 RH-15?:2 RH-16:1
db745675bd	v1.91.33.0 feat(ship): measurement bar with explained-red and pooled dec	26	18	18	0	16	0	RH-15?:1
28f1385eac	v1.91.32.0 fix: Codex works from zsh, free text never enters shell comma	120	65	50	15	155	0	RH-15:1 RH-16:1
7697fb2981	fix(cso): install Rails gems on one thread so exports stop failing (v1.9	2	0	0	0	7	0	-
55f8151fbe	v1.91.30.0 fix: honest eval failure reports, Bun floor with safe auto-up	414	219	50	169	103	0	RH-1:2 RH-13:4
5885157a90	v1.91.29.0 fix(cso): native runtime qualification passes for Node, Bun,	21	12	12	0	14	0	-
c285d88b90	Promote qualified CSO scanner catalog (#3042)	0	0	0	0	1	0	-
10315cf44d	v1.91.27.0 fix(cso): runtime staging can finish and a qualified scanner	11	7	7	0	7	0	-
857466ff8b	v1.91.25.0 feat(cso): private two-architecture release scoring + self-ch	55	30	30	0	19	0	RH-15?:1
55fbe77195	v1.91.24.0 test: audit wave — PR evals keep finished work, honest ceilin	761	637	50	587	66	0	RH-1:14 RH-14:2 RH-15:1 RH-15?:8 RH-16:1
f97aefe654	v1.91.22.0 feat(cso): reviewed, attested scanner build inputs + fixes so	5	3	3	0	10	0	-
2db0b3adc8	v1.91.19.0 fix: severe fix wave (truthful gates, no lost memory, every h	577	317	50	267	219	0	RH-1:1 RH-12:1 RH-16:11
4015c2870b	v1.91.18.0 feat: Opus 5.5 cleanup follow-ups — scoped transcript consent	423	110	50	60	82	0	RH-1:2 RH-16:1
fe6d1ae62a	v1.91.17.0 fix: mvanhorn wave (checkpoint provenance, never-overwrite de	24	8	8	0	37	0	RH-16:3
ac20ef1e13	v1.91.16.0 fix: safe installs and upgrades, skills start everywhere, tru	420	288	50	238	228	0	RH-15:2 RH-15?:2 RH-16:1
74512c20a7	v1.91.15.0 v1.91.15.0: Opus 5.5 prompt cleanup, behavior-level skill tes	538	451	50	401	186	0	RH-1:8
f30b7b788a	v1.91.13.0 fix: severe bug wave (autoplan guard, land-and-deploy CI gate	115	66	50	16	56	0	RH-16:1
7fca42ad8b	v1.91.12.0 v1.91.12.0: audit fix wave, ~11-minute paid eval lanes, eval	712	516	11	505	109	0	RH-1:4 RH-12:1 RH-13:1 RH-15:7 RH-15?:7 RH-16:23
df89475b17	v1.91.11.0 refactor: one state-root rule, browse route table, shared sha	310	223	16	207	236	0	RH-1:5 RH-13:3 RH-15:1 RH-16:3
96764e80a6	v1.91.9.0 feat: test value bar in plan-eng-review, review, qa and ship,	45	33	33	0	37	0	-
943105f109	v1.91.8.0 test: smaller suite, derived paid selection, retired never-gre	905	858	37	821	30	3	RH-1:344 RH-15:1 RH-15?:1
dcaea52800	v1.91.7.0 feat: add functional QA and pre-publication docs checks (#2983	670	492	50	442	138	0	RH-1:7 RH-15:4 RH-16:1
65bfb0ce49	v1.91.6.0 feat: balance PR eval slices by recorded duration without weak	2	0	0	0	7	0	-
080a655791	v1.91.5.0 feat: balanced free-suite shards, test:ubicloud, and faster PR	36	23	23	0	12	0	RH-1:1 RH-15:1 RH-15?:1
d2a0bbcf4c	v1.91.4.0 fix: Windows Opera cookie import wave (#2980, refs #2957) (#29	12	5	5	0	14	0	-
01593aa67c	v1.91.2.0 fix: consolidate gstack reliability wave (#2959)	269	166	50	116	97	0	RH-1:3 RH-15:5 RH-15?:1 RH-16:4
2a113ae7e6	v1.91.1.0 fix: harden Impeccable plugin discovery (#2978)	27	3	3	0	6	0	-
7b534d3e90	v1.90.2.0 perf: halve local free-suite time and preserve coverage (#2972	99	56	50	6	21	0	RH-1:1
a84b0b5b6d	v1.90.0.0 feat: make browser cookie imports explicit and safe (#2964)	160	74	10	64	35	0	RH-12:3 RH-15:3 RH-15?:1 RH-16:11
730a1017d1	v1.89.1.0 fix: remove continuous checkpoint commits (#2970)	98	72	50	22	69	1	RH-1:1
06ed920a97	v1.89.0.0 feat: add shared-code extraction audit (#2925)	267	189	50	139	89	3	RH-1:2 RH-15:1
b9706f3635	v1.88.1.0 fix: harden credential boundaries and owned state (#2942)	59	42	42	0	12	0	RH-1:1 RH-12:1 RH-16:8
636175d349	v1.87.6.0 fix: make checks reliable and everyday validation faster (#289	1276	618	50	568	154	2	RH-1:12 RH-14:1 RH-15:6 RH-15?:3 RH-16:4
35dd014c58	v1.87.5.0 perf: remove idle waits from tests and CI planning (#2897)	77	47	47	0	12	0	-
a6b3a57512	v1.87.4.0 fix: preserve health failures and disclose coverage (#2882)	15	6	6	0	6	0	-
85b8c038fc	v1.87.3.0 fix: bind review evidence to the reviewed tree (#2875)	32	19	19	0	29	0	-

Scanned: 50 of 50 first-parent commits on origin/main; 0 errored. Listed per commit: median 40.5, p75 267, max 1276; eligible median 24.5, p75 126, max 858; partial (read cap hit) in 20 commit(s). Tags: RH-1: 411 hunk(s) in 18 commit(s); RH-12: 6 hunk(s) in 4 commit(s); RH-13: 8 hunk(s) in 3 commit(s); RH-14: 3 hunk(s) in 2 commit(s); RH-15: 35 hunk(s) in 14 commit(s); RH-16: 80 hunk(s) in 19 commit(s); RH-15?: 30 hunk(s) in 13 commit(s).