Publish seeded web-research follow-up with candidate qualification gates - #751

Scope

Publish the user's seeded web-research sweep landscape-sweep-webresearch-20261005-ns2604 (workflow wf_5d8a2a30-9b5) through the existing landscape harness steps 9–12. Sixteen formal proposals yielded four targeted candidates, IPI-proxy's preserved not_adopted disposition, and eleven fit refutations. The PR retains the dated manifest, source reviews, original converted evidence/usage/seeds/scope and an append-only saturation record; current selections, host acceptance and installation plans are unchanged.

• Base commit: 2d849ba1fe1b879144cfa470f0514d3ac3b8e487; original frozen publication inputs/tests started at e7d94e7bcc074edba093994c8853e59b650ce7e0.
• Lane: lane:foundation; this PR stays draft for the command center's Claude read.
• Owned paths: unique sweep artifact directories, catalogs/sota-convergence/manifest-20261005-webresearch-ns2604.json, the dated decision and index paragraph; native append/generator projections and manifests/evidence.json in the last commit.

SOTA sources

• Native publication procedure: native-agent-stack@e7d94e7b, landscape-sweep README:638, saturation recipe:89, hot-file protocol:94.
• Exa MCP@f3d71fb6 README, Trafilatura@6c1977a0 README, LettuceDetect@2096ed28 README, CDX Toolkit@e87b1204 README, IPI-proxy@272a61a7 README; native generated provenance reviews retain these exact commits.
• Benchmark scope correction: web-search-for-coding-agents@a4fc6411 judge.py:89, :195; Exa search type@f3d71fb6:68.

Evidence-class table

Local commands run

python3 tools/sota-convergence/build_manifest.py --work-dir <private-frozen-inputs>/work --lanes <retained>/out/lanes.json --checkout-root <original-source-checkout> --checkout-root <original-lane-checkout> --out <private>/manifest-20261005-webresearch-ns2604.json --checked-at 2026-10-05 --id sota-convergence-20261005-webresearch-ns2604
exit 0
python3 tools/sota-convergence/landscape-sweep/source_reviews.py --survivors <retained>/out/survivors.json --out evidence/artifacts/landscape-sweep-webresearch-20261005-ns2604 --lane landscape-sweep-webresearch-20261005-ns2604
exit 0; five reviews
host_receipts.register_file(...) for each owned artifact and changed registered source
exit 0; repeated from main's registry after clean rebase
python3 tools/sota-convergence/landscape-sweep/make_result.py --layers <retained>/out/layers.json --reviews <private>/out/reviews.json --sweep-id landscape-sweep-webresearch-20261005-ns2604 --returns-ref evidence/artifacts/landscape-sweep-webresearch-20261005-ns2604/returns.json --usage-ref evidence/artifacts/landscape-sweep-webresearch-20261005-ns2604-attempts/child-usage-wf_5d8a2a30-9b5.json --manifest-ref catalogs/sota-convergence/manifest-20261005-webresearch-ns2604.json --work-dir <private-frozen-inputs>
exit 0
python3 scripts/saturation_ledger.py --append <private>/RESULT.json
exit 0; main's existing sweep prefix preserved exactly
python3 scripts/saturation_ledger.py --check --base <base-commit>
exit 0; six records, intact chain/bindings
TMPDIR=<owned-cache>/tmp python3 -m unittest tests.test_saturation_ledger tests.test_landscape_sweep_harness
exit 0 on original base: 388 tests, 4 skipped, 96.784s; exit 0 on final base: 388 tests, 4 skipped, 103.869s
python3 scripts/component_matrix.py --write / --check
exit 0; 32 rows, zero flip-rule violations
python3 scripts/new_host_grand_list.py --write / --check
exit 0; 32 layers, 66 winners
python3 scripts/validate.py
exit 0 before source commit; post-rebase exit 1 for an unregistered final ledger hash; native registration corrected it; final exit 0: 69 components, 10,221 hashed files, 4 profiles, 207 receipts
git diff --check
exit 0

Exact private argv, failures and final integrated results are recorded in the lane status. The failed post-rebase check is retained: the final appended ledger needed re-registration before validation. No test or validator was weakened. Read-only locator/syntax mistakes were corrected before publication; original returned evidence was never rewritten.

Decision record

docs/decisions/2026-10-05-seeded-web-research-landscape.md names alternatives, source pins, original votes and each measurement that would overturn the disposition. Adoption still requires that measured comparison and native organic-use qualification; this PR authorizes neither a provider flip nor a runtime install.

Scope corrections are explicit: Browser-use/Playwright MCP are discovery-stage dropped entries, not formal refutations. The original critic covered the first eight before the eight-proposal follow-up. IPI-proxy remains rejected under this wave's maintenance rule. The benchmark judge's URL-or-token rule does not establish strict citation support; a repository pin does not freeze the hosted MCP service. CDX equality must account for published redactions/HTML-to-text transforms. The configured promptfoo assertion question remains attributed reviewer evidence rather than a current-client absence claim.

Host evidence

No evidence/hosts/ change or platform-status promotion. The co-op owns cross-family draft review, any later qualification/activation and landing; this lane does not merge.

Checklist

• Existing native harness commands used; no new runner, installer or validator executable.
• No GitHub Actions changes.
• Registry and shared append/generator outputs committed last from main's copies.
• No credential value/file read or copied; no raw native transcript, active client configuration or private host path published.
• No new paid service, subscription, install, configuration or plan row.
• Peer-owned paths/checkouts preserved; the original lane draft was untouched.
• Completeness findings feed the next web-research sweep: missed/dropped surfaces, native browser/platform exposure, point-in-time archive leakage, PDF/table fidelity, false-accept controls and browser recovery.