[crazydemo]

Description

CBTS currently tries to infer runtime consumers through callers, decorators, cached values, reflection and shared pytest state. These bounds can force full CI even when the changed code has directly recorded coverage. Full CI also does not establish whether newly introduced paths have adequate assertions.

Map verified Python patches to the union of directly changed symbols in the old and new source images, then query the coverage DB by (file, qualname). Missing file/function records contribute no inferred tests. Deleted functions use their old symbols. Module/class/definition-time changes additionally require a registered import-sanity inventory. For example, a cached method body change selects its recorded tests without traversing cache-hit callers; changing a module binding queries <module> and retains import sanity without inferring every reader.

Preserve PR-added/modified tests, explicit rule contributions, unknown test entries, incomplete collection records, shared-block protections, CPU obligations and the existing multi-GPU gate. Missing or unusable source/patch data, unsupported residual files and incompatible coverage baselines still prevent narrowing. The legacy --no-data-policy values remain accepted; all now use the lookup-only behavior and the default becomes ignore.

Add cold subprocess imports and a unified 28-selector sanity inventory, with CPU/package-sanity registration and the missing H100 registration. Retain registered test consumers for explicitly mapped test infrastructure. Update selection diagnostics, documentation and regression coverage.

Test Coverage

• Complete CBTS unit suite: 1027 passed, 6 skipped, 2 subtests passed. Includes old/new diff attribution, missing coverage rows, deletions, import obligations, test retention and infrastructure rules.
• Repository pre-commit and DCO hooks passed. The committed branch merges cleanly with upstream main 5f66e8a10e in a local merge-tree check.
• Fixed historical replay: the same 24 PR snapshots and DB3063 (2d5b1fb68fee16b49b43f501d9165af23fb83c33), pre-merge mode, multi-GPU gate closed. Both policies use the same complete sanity registrations with matching CPU/non-CPU marker eligibility. DB SHA256: b37654fac2fe0fe102ba1cda82f52b6208508258590c59f75b0344ea1591a698.

Fixed-cohort metric	Previous policy	Lookup-only policy
Selector hits	14/24	22/24
Hits after baseline compatibility	14/24 (58.3%)	21/24 (87.5%)
Remaining fallback	10	3
Selected YAML entries after fallback	26,186 / 40,777	25,204 / 40,777
Effective weighted skip rate	35.78%	38.19%

Seven additional compatible hits: #20903, #21204, #21220, #21463, #21192, #21466 and #21121. No previously compatible hit was lost. Remaining fallback: #21280 has C++ residual changes; #20104 has CUDA residual changes; #21225 has a patch conflict with the coverage baseline in mega_moe_cute_dsl.py.

The replay checked 1,008 required sanity-retention obligations and 8,274 direct-coverage or untrusted-entry obligations; 9 replay-alignment helper tests passed, with no replay errors. Counts are weighted YAML entries, not expanded pytest cases or measured GPU savings. The replay executes the selector, not historical product builds. The real native/CUDA cold-import inventory has not been executed at this PR head; passing retained or full CI does not prove coverage of new paths.

Fixed replay cohort and per-PR outcomes

PR	Frozen head	Previous scope	Lookup-only scope	Baseline compatible	Selected / total entries
#21777	d9a177c57a	coverage	coverage	yes	1597 / 1767
#20903	042776e170	fallback	coverage	yes	1647 / 1767
#21704	3b4bbce3d1	testsonly	testsonly	yes	128 / 1768
#21631	17b03c69d7	coverage	coverage	yes	1430 / 1775
#21204	3b74af0c2a	fallback	coverage	yes	1404 / 1769
#21874	31903dc655	noop	noop	yes	0 / 1775
#21692	4f9a5bda8e	testdefonly	testdefonly	yes	44 / 1768
#21220	9ae1819125	fallback	coverage	yes	1520 / 1611
#21280	5208f1284d	fallback	fallback	no	1635 / 1635
#21463	004f6cc4ec	fallback	coverage	yes	1628 / 1724
#20104	b7dae02645	fallback	fallback	no	1281 / 1281
#21144	21f0dbd18b	coverage	coverage	yes	1661 / 1764
#21192	448febe2a6	fallback	coverage	yes	1710 / 1715
#21249	7017065e01	testsonly	testsonly	yes	118 / 1720
#20918	d246bdaa61	coverage	coverage	yes	1386 / 1722
#20089	fc65496d9f	coverage	coverage	yes	1392 / 1617
#21558	4760c852f0	testsonly	testsonly	yes	259 / 1718
#21466	63a7f7d13a	fallback	coverage	yes	1557 / 1712
#21199	62902cec2a	coverage	coverage	yes	1551 / 1716
#21121	8a28b0216d	fallback	coverage	yes	1541 / 1700
#21465	bc77820e71	testsonly	testsonly	yes	2 / 1712
#21302	4861cb1059	noop	noop	yes	0 / 1635
#21225	90e95a4c0d	fallback	coverage	no	1559 / 1706
#21387	e9799f0706	waiveonly	waiveonly	yes	7 / 1700

The table shows raw selector counts; effective totals above restore the full entry count for baseline-incompatible selections.

PR Checklist

• Reviewed coding guidelines and updated selection documentation.
• Added regression coverage and CI registration for the new sanity test.
• No public product API, dependency, ownership or model-architecture changes.