Read the newest sweep manifest in the new-host grand list - #1000

Scope

The new-host grand list reads the newest committed sweep through the existing component-matrix selector, replacing its September 23 constant. Its currency flag now compares the displayed winner pin with that sweep's upstream release using the producer's own comparison API, correcting 14 falsely current rows and the distinct-component summary from 20 to 31. The preceding guard repair validates the selected document at the report's new join boundary: incomplete sections, missing report layers, malformed consumed component fields, and a future UTC checked_at stop the build with the offending path before output or registration. It never falls back to an older sweep.

• Base: 65cf2730068429a1c48c8c18fb24aadef8120817.
• Head: c7ab11920d5bff32f50ce43a8ef3f763db0b1e04, ONE CC-adjudicated correction of 74721739be522a1bcfdd972c6544c411c813a5b5, after the earlier guard repair of adca78fb4e6f14408de815a71c6435c932882abd.
• Lane: lane:foundation.
• Original owned paths: the grand-list script, tests, generated JSON/Markdown and their existing evidence bindings. The pin correction changes the script, its fixture test, both generated reports and their three existing evidence bindings. The shared selector, native producer and bootstrap pin files remain byte-identical.

The maintained selector still chooses the greatest parsed checked_at and uses the later filename for a tie. The report checks the shape of fields it consumes, including the manifest's coverage of report layers. Null pin comparisons remain unknown; candidate records retain their native repository/disposition shape, and the existing five unjoined winners remain explicit. This does not claim whole-document sweep-schema validation.

SOTA sources

• [mechanism] 65cf273:scripts/component_matrix.py:L659-L699 provides the existing JSON/date selection and duplicate-layer refusal. Its Git blob is ecc083fff39c87853fc675b46729cff766e2d273, identical at reviewed adca78f. No second selector is introduced.
• [mechanism] 65cf273:tools/sota-convergence/build_manifest.py:L1818-L1862 assembles the foundation components and trading entries with id, repository, pin and upstream/comparison fields; L1971-L1978 defines complete layer coverage; L2021-L2024 preserves null for uncomputed comparisons. Git blob e3afbcb445d2719a44b6601c1993cd7a6b0649ef is identical at reviewed adca78f. The consumer guard follows this maintained producer contract.
• [mechanism] python/cpython v3.13.16, locked commit cbc944f4bc59639a444dd971c737788ba2283a91, Doc/library/datetime.rst:L925-L939 and L951-L958 documents timezone-aware datetime.now(timezone.utc). The installed mirror was read at that pin; the consumer compares the sweep's parsed date with the current UTC date, using a fixed clock in regressions.
• [reference] 65cf273:catalogs/saturation/ledger.json:L10918-L10922 records the completed October 9 sweep; docs/lanes.md:L94-L116 specifies report regeneration and registration of files already listed on main; .github/workflows/validate.yml:L151 checks the grand list.

Evidence-class table

The table and original command block below record the preceding 7472173 guard repair. The adjudicated displayed-pin correction has separate current validation below.

Local commands run at preceding 7472173 guard repair

python3 -E -s -B -m unittest tests.test_new_host_grand_list tests.test_component_matrix tests.test_evidence_manifest
112 tests, no skips; exit 0
python3 -E -s -B scripts/new_host_grand_list.py --write
32 layers / 66 winners; both reports byte-identical to reviewed head
python3 -E -s -B scripts/new_host_grand_list.py --check
exit 0
python3 -E -s -B scripts/component_matrix.py --check
32 rows; exit 0
python3 -E -s -B scripts/evidence_manifest.py --check
11,472 files; exit 0
nice -n 10 ionice -c 2 -n 7 python3 -E -s -B scripts/validate.py
70 components / 11,472 hashed files / four profiles / 239 receipts; exit 0
git diff --check
exit 0
gitleaks git --redact=100 --log-opts='65cf2730068429a1c48c8c18fb24aadef8120817..74721739be522a1bcfdd972c6544c411c813a5b5'
two PR-range commits, no leaks; exit 0

Decision record

Both reproduced GPT P2s are blockers on the new caller; their underlying shared-selector omissions are not repaired across unrelated consumers. The adopted selector, date ordering and filename tie behavior remain the recorded mechanism. No new library, candidate selection, verdict or catalog pin is introduced.

The earlier 12-to-20 sweep-pin count did not describe the displayed winners. The CC's newly reproduced real-data failure supersedes its prior P3 classification: 14 joined rows for 11 components were falsely marked current. The correction reuses the producer's own classify_pin and pin_comparison_fields APIs for each displayed winner pin against the joined upstream release. The true distinct-component count is 31, and the generated page now discloses this exact comparison basis beside its separate bootstrap-pin note.

Adjudicated displayed-pin correction

Primary mechanism: 65cf273:tools/sota-convergence/build_manifest.py:L39-L48 and L147-L176 owns version, commit-tracking and OS-package exclusion behavior; L1366-L1372 owns null/not-compared output. The file's Git blob e3afbcb445d2719a44b6601c1993cd7a6b0649ef is unchanged here. The report loads that existing module by exact code path using the producer's own supported importlib pattern at L113-L123; fixture input ROOT does not select executable code. No second comparison rule is written.

Fail-first at exact 7472173: the new fixture deliberately sets the sweep's pin to 2.0.0/current while the displayed winner varies. One method gives two intended assertion failures and zero errors: displayed 1.0.0 should be behind 2.0.0; a displayed commit-tracking pin should remain null/not compared. The equal-version control already passes. After repair all cases pass, including the rendered behind note and summary.

On committed data the correction changes exactly 14 row flags from false to true and the summary from 20 to 31 distinct components; no other JSON property changes. There are still 66 winners, 61 joined and five unjoined. Restored components are agent-browser, ccusage, claude-code, headroom, markitdown, mcporter, opentelemetry-collector-contrib, prometheus, rtk, sandbox-runtime and worktrunk. Both JSON and Markdown are regenerated; the page's behind notes now describe the displayed pins.

Current gates: 270 tests without skips (grand-list 44, component-matrix 64, evidence five and the maintained producer's SOTA-convergence module 157), FULL 70 components / 11,472 hashed files / four profiles / 239 receipts, report/matrix/evidence checks and diff check all pass. Native registration updates only the test and two report bindings. The actual three-commit PR range scans with 100% redaction and no leaks.

Follow-ups

• P3: add direct shared-selector tests whose filename order disagrees with checked_at and which exercise the filename tie rule; these are currently independently probed.

• Wider scope: bring the component matrix and other shared-selector consumers under an agreed sweep-admission/whole-document schema contract; the report guard covers its own consumed fields only.

• CC-adjudicated follow-up 4242517442: the producer selects a prerelease stream for the sweep pin's major, which can differ from the displayed pin's major. Zero committed rows are affected; carrying release-list evidence or selecting the displayed pin's stream remains a producer/consumer follow-up.

• CC-adjudicated follow-up 4242517447: equivalent non-canonical GitHub URLs can join but fail the producer classifier's URL recognition. Zero committed rows are affected; align repository normalization with the native producer.

• CC-adjudicated follow-up 4242517453: custom --os-package-ids cannot be recovered from a sweep manifest that omits generation options. Zero committed rows are affected; the producer must record its options before consumers can preserve a custom exclusion policy.

• CC-adjudicated producer follow-up 4242517457: Promptfoo is compared with code-scan-action-0.2.2 from the producer's upstream stream selection. These values and the flag are unchanged since 74721739; correct the producer's stream selection separately.

• Cosmetic P3: align the indentation of the two added page-note strings at scripts/new_host_grand_list.py:572-573 with their neighbours.

• Astra P3: improve the shared selector's malformed-layer diagnostic so it includes the offending manifest path.

• Astra P3: refresh docs/decisions/2026-10-05-dagu-mise-pin-move.md:128, whose live-row citation still targets the old scripts/new_host_grand_list.py:44 constant line.

Host evidence

No host receipt, installation, service, broker, model or provider operation. Synthetic fixtures and local repository gates do not establish hosted acceptance. The CC owns review and landing.

Checklist

• Both in-scope blockers reproduced on the exact reviewed head and repaired in one forward commit.
• Permanent fail-first tests protect refusal, outputs and registration.
• Both JSON and Markdown reports were regenerated; three existing evidence bindings changed (the two reports and the test). The shared selector and native producer retain their reviewed bytes.
• Touched modules, FULL validation, evidence and diff checks pass.
• PR-range-only gitleaks uses 100% redaction; no credential or settings read.
• Confirmation reviews, actual-head hosted CI and CC landing.