foundation: register the pending current model-routes scope - #1179

Scope

Register foundation/model-routes as an explicit pending-empty scope so the documented two-family landscape sweep can address it. The catalog, taxonomy, research and quality coverage use the CC's exact 19:19Z sentence; native generated views retain status unswept, zero winners, no terminal state and no wave.

• Base commit: 8053f14e7e0817d2502dff2bb284bfb7e1b0152f (current main).
• Review-wave fold: 92c95deff39b8106534407e0b75f5c94f53504bd, one commit directly after reviewed head eedd595e06833706d6d7b5b1b3f4855786ee19ce; no rebase or history rewrite.
• Gap-ledger P2 fold: 2bda440abd751d9b8e9cb5bbc747a1071d1a5f72, one commit directly after 92c95deff39b8106534407e0b75f5c94f53504bd; the CC FULL's three P3s are recorded in Follow-ups below.
• Lane: lane:foundation.
• Predecessor: closed Register the pending model routing layer for its covering sweep #1126, held snapshot 0f9e90b1a31f10c213857b36a6d4aa5d5ee4290a. Rules prerequisite: Prepare rules for unrun pending scope registrations #1130, landed as f165498aa6340860ef52de3a720ca2981eced062. This PR rebuilds registration data on main and carries no stale rule code from the predecessor.
• Owned paths: foundation taxonomy and landscape row, matching research/quality/surface coverage, native matrix/final-stack/new-host outputs, dependent handbook outputs and receipt, active disposition locators, historical audit integration, dated decision, regressions and controls, evidence registrations.

The exact requirement is:

Scope: current model ids and routes for the Anthropic API (vendor docs and the claude-api skill), OpenAI through OmniRoute, and Meta's Muse through the muse-code provider (Wonder Full; the Muse fence rules hold).

SOTA sources

• Pending-empty admission at scripts/landscape.py:67 and build_landscape:1284: the landed registration schema and full catalog/coverage join; no validator change here.
• Foundation taxonomy at scripts/validate_foundation.py:72 and wave classification at build_verdicts.py:164: exact matching scope, registration-only capability boundary and no frozen-wave membership.
• #1130 decision, layer-resolution protocol and carried angle record: absent verdict/sweep is a named gap; carried source-review candidates are not selected by registration.
• Muse file fence and named toolchain grants: current fence rules remain binding; no Muse invocation or isolation change.
• python/cpython v3.13.16, Lib/unittest/case.py and loader.py: the repository's existing regression harness.
• Publication consumers at reviewed commit eedd595e06833706d6d7b5b1b3f4855786ee19ce: build_ecosystem.py:293 requires exact taxonomy/surface coverage and supports empty bindings; build_new_wsl_handbook.py:187 binds input bytes; audit_reports.py:26 preserves canonical decisions and dated synthesis pointers; shared unrun admission distinguishes new pending registrations from selected, sealed or previously swept layers.

Evidence-class table

Initial publication checks at eedd595

All checks used the existing idle-class scope wrapper.

python3 -m unittest tests.test_model_routes_registration
before: exit 1, 3 tests/3 failures, missing explicit registration (0 != 1)
after: exit 0, 3 tests, no failures or errors

python3 -m unittest tests.test_registration_rules tests.test_pending_catalog_registration tests.test_layer_verdicts tests.test_record_verdicts tests.test_model_routes_registration
184 tests passed, zero failures/errors/skips (native TestLoader/TextTestRunner; fixture output captured)

python3 scripts/validate_foundation.py --root . --json
python3 scripts/landscape.py
python3 tools/sota-convergence/build_verdicts.py --check
python3 scripts/component_matrix.py --check
python3 scripts/new_host_grand_list.py --check
python3 scripts/final_stack.py --check
all passed; 21 taxonomy scopes, 33 catalog layers; frozen wave retains 20 foundation + 12 trading rows

python3 scripts/verdict_review_gate.py --base origin/main --json
passed; trust_paths_changed=[], rule_inputs_changed=[], waves_changed=false, violations=[]

python3 scripts/validate.py
passed: 71 components, 12142 hashed files, 4 profiles, 240 receipts

All 12 changed/new published-file registrations match. Whitespace checks and installed gitleaks with redaction pass; staged scan found zero leaks. No new scripts or source-policy grants are needed. The normal pre-push privacy/history, registry and digest checks run without override.

Normal pre-push checks passed: 13,001 files scanned with zero privacy matches, three registry inventory tests passed, and the published-digest privacy guard passed. The no-reply author identity was checked before the normal push. Main stayed at the stated base through publication; no rebase was needed.

Decision record

docs/decisions/2026-10-15-model-routes-catalog-registration.md.
evidence/artifacts/model-routes-registration-20261015/controls.json retains fail-first and pass-after counts.

Review-wave corrections

The initial six catalog checks omitted additional publication consumers. The local review and GPT full read at eedd595e identified six consequences of adding the 21st foundation scope. Private reads are cited as pr1179-eedd595e/review_output.json and GPT-VERDICT-eedd595e0.md, without digests of unpublished bytes.

1. Add the supported empty model-routes surface/runbook binding. The new native build_data regression and ecosystem publication check first failed with the exact layer-coverage error; four registration tests and the ecosystem check now pass.
2. Repin active sandbox disposition locators after searching moved citations. The eight citation controls first had two added failures, then passed. Fixed-pin and dated historical citations retain their historical coordinates.
3. Regenerate the new-WSL handbook with its owning builder and rebind its current output receipt, retaining previous bindings. Only the research-state input hash changes in the generated outputs; its generator, profile, selections and 37-layer edition remain intact. The stale-output test first failed, then passed.
4. Add an explicitly pending, no-winner grand-catalog summary row. The prior 20/21 coverage failure is corrected while exact winner/decision checks remain for all recorded rows.
5. Name the historical blind audit's 2026-09-21, 32-layer scope and exclude only registrations admitted by the shared predicate. Historical synthesis and handbook bytes remain unchanged. Added controls refuse winner, sealed-run or prior-sweep exemption misuse; these controls first failed and now pass.
6. Update architecture counts to 21 foundation/12 trading layers, preserving exact identity checks and adding pending/no-winner/no-wave/no-FINAL assertions. Both affected count assertions failed before the correction and now pass.

The initial targeted correction run passed 19 tests. The broader affected-suite run executed 496 tests with one failure, no errors and no skips: the publication control caught stale evidence registrations for the edits. After refreshing all 13 edited-file bindings through scripts.host_receipts.register_file, that exact test passed in 62.735 seconds. This is 495 passing first-run tests plus the resolved one-test rerun, rather than a claim that the whole selection ran again. After release, the final fail-first recheck reproduced all six mechanisms on a clean reviewed-head export (24 tests, 16 assertion failures, one surface-join error, zero skips); the final-tree targeted selection then passed all 24 tests with zero failures, errors or skips. Final python3 scripts/validate.py passes: 71 components, 12,144 hashed files, four profiles and 240 receipts. All 13 edited-file registrations match and whitespace checks pass. Native ecosystem/new-WSL/blind-audit/frozen-wave checks and the base-trusted gate pass; the gate reports no trust-path, rule-input or wave changes.

The fold's normal pre-push checks also passed: 13,001 committed files scanned with zero privacy matches, three registry tests passed, and the published-digest privacy guard passed. No-reply author and committer identities were verified without printing them. The push was a fast-forward from the exact expected reviewed head, with no rebase or history rewrite. Staged gitleaks scanned 36,307 bytes with zero leaks.

Dated gap-ledger review correction

The GPT delta at 92c95deff dispositions the prior six blockers as closed. Its historical gap-ledger coverage correction now names the 2026-09-22/32-row scope and exempts only shared, history-aware admitted unrun registrations. The historical ledger and all source/receipt bindings stay byte-identical. Primary sources at 92c95deff39b8106534407e0b75f5c94f53504bd: scripts/landscape.py:67-95, tests/test_gap_resolution_ledger.py:22-28, catalogs/landscape/gap-resolution-20260922.json:3-11, and blueprints/blind-catalog-convergence/audit_reports.py:43-46. Original seven tests first had one failure; nine tests with added winner/sealed/prior-sweep controls first had four failures. After correction, 21 gap-ledger/blind-audit/registration tests pass with no failures, errors or skips. Unmarked extra rows and missing dated rows still fail coverage. The CC FULL at that same reviewed head, MICRO-full-92c95def.md, passes; both verdicts are received, and the co-op authorized one P2 fold commit. Review records are cited by name only.

The review's environment-limited execution counts (head: 92 modules/4,602 tests; supplied merged overlay: 100 modules/4,832 tests) are counts, not blanket passing claims. Its genuine-worktree receipt reports 13 modules/1,017 git-dependent tests at base and head with rc 0 and disclosed skips; that does not establish later-merge results. Missing source captures and historical pins remain gaps.

Local P2 verification also passes full scripts/validate.py (71 components, 12,144 hashed files, four profiles, 240 receipts) and the base-trusted gate (no trust-path/rule-input/wave changes or violations). All three edited-file evidence bindings match. This fold changes only the gap-ledger test, decision, controls and evidence registry. The final recheck again passed 21 targeted tests, and the pre-correction row collection again produced the expected four failures across nine tests. Staged redacted gitleaks scanned 11,266 bytes with zero leaks. No-reply author/committer and expected remote-head lease were checked; the normal fast-forward push passed the 13,001-file privacy scan, three registry tests and the published-digest guard, without overrides.

Follow-ups

The CC FULL's three P3s, retained by GPT-VERDICT-2bda440ab.md, remain follow-ups. The round-three disposition is recorded by name in adjudication-pr1179-2bda440a-20261015.md:

• The later comparison's field list in the decision at reviewed lines 51–54 remains a PROPOSAL. Its dimensions need pinned primary-source evidence before they become comparison requirements; the registration establishes no route default.
• The component evidence matrix's newest verdict checked_at summary includes the registration date (docs/component-evidence-matrix.md:3,174 and its JSON at reviewed line 9409). A follow-up in the owning generator should distinguish registration dates from completed-verdict dates; this timestamp does not establish a completed model-routes verdict.
• The new-host grand list's generic incumbent wording at reviewed line 137 also reaches model-routes, which has no incumbent. A follow-up in the owning generator should distinguish pending registrations from retained selections; no incumbent is asserted for model-routes here.

The inherited owner-chat P1 identified by GPT-VERDICT-2bda440ab.md belongs to the separate scrub PR #1161 under the CC adjudication above. #1179 adds no paraphrase or synthetic-quote work. Its head remains held at 2bda440abd751d9b8e9cb5bbc747a1071d1a5f72 until #1191 and then #1161 land and the co-op relays #1161 MERGED. Then #1179 rebases onto the resulting main and regenerates the handbook through python3 scripts/build_new_wsl_handbook.py --write, with only forced binding refreshes. The resulting delta requires one GPT read and the CC delta before landing; no history rewrite of main is involved.

• Pinned-source captures requested by the reviewer remain a named evidence gap; source links alone are not captured upstream bytes or measured route capability.
• Environment-limited reviewer results remain gaps. The local affected-suite run's actual skips are reported with its counts rather than inferred from the review environment.
• The reviewer reproduced 15 inherited upstream-surface-watch failures at both base and head. The eight affected disposition citation tests pass after repinning; that does not claim the entire module passes.
• The review retains a moved-line inventory of 210 citing lines in 25 files. Active tested disposition references are corrected; dated evidence retains its pinned coordinates.
• New practice or tool claims require upstream sources at a pin and measured agreement; an unestablished proposal remains a PROPOSAL. Registration selects no route or practice winner.

Named gaps and next inputs

• No ledger-countable two-family sweep has run for this scope.
• The two single-family initial sweeps are separate: the GPT half (Codex lane mage, open foundation: retain GPT-only initial model-routes sweep #1177) and the Claude half (co-op under CC's 19:52Z option (A)). A later separate PR records them as PRIOR evidence for the documented two-family sweep. They are never this PR's verdict and never FINAL.
• Each keyed measurement waits for CC approval. This PR makes no keyed call, install, credential read, provider/account mutation or route activation.
• The CC reads this registration PR before the co-op's landscape-sweep --slots 2; later covering-sweep publication and candidate/taxonomy promotion remain separately owned admission/lifecycle work. No guard is relaxed here.

Host evidence

No host receipt or platform-status change.

Checklist

• Empty candidates, winners and alternatives; pending unsealed lanes; named gaps; no terminal state or wave claim.
• No verdict-admission rule, frozen wave, gate, dependency, paid service or account-state change; historical audit integration uses the existing shared admission predicate.
• Private source records cited by filename without unpublished-byte digests; no owner quotations or private identifiers.
• Peer-owned branches and artifacts preserved.
• CC FULL at the reviewed 92c95deff head passed; its three P3s are retained above.
• Relock the OpenHands recipe onto mcp 1.30.0 (GHSA-84m7-p3x7-pcfv) #1191 then Scrub owner quotations from the new-WSL records: paraphrase the canonical inputs, regenerate, and carry a dated paraphrase in the relay contract #1161 land, co-op relays #1161 MERGED, and the authorized rebase/regenerate-only delta passes GPT and CC reads plus green CI before landing; any later sweep keeps its own authorization and readiness requirements.