Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@config/sota_v2_models.json`:
- Around line 161-178: Remove the expiring route entry identified by id
openrouter-tencent-hy3-free-novita, or replace it with a stable, smoke-tested
route that remains valid beyond the freeze. Do not retain this route unless the
publication process is explicitly bounded before its catalog_expiration_date.
- Around line 14-15: The frozen model panel in the selection_policy and models
registry must contain twelve models: five proprietary and seven open-weight.
Update the model entries and associated required_smokes coverage so the declared
cohort, smoke completeness, cost planning, and leaderboard eligibility
consistently use the required five-plus-seven composition.

In `@docs/PUBLISH_READINESS.md`:
- Around line 13-17: Synchronize the frozen ten-model phase-one state across
docs/PUBLISH_READINESS.md lines 13-17, 30-32, 136-165, and 410, marking fresh
4,096-token smokes, verification, cost-plan, and lane-freeze work as complete
and removing provisional or retired-plan wording. Update
docs/output_budget_sweep.md lines 3-7 to call the current policy frozen, and
lines 62-65 to remove excluded Kimi K3 from the frozen lane’s reasoning
configuration.

In `@results/analysis/output-budget-cost-estimate.json`:
- Around line 127-136: Regenerate the runtime evidence in the output-budget cost
estimate from the accepted 4,096-token smoke observations, replacing the legacy
1,024-token source and updating the observation timestamp. Populate
observed_api_seconds_per_decision_by_model for all ten models in the accepted
ledger, rerun the estimator, and update the runtime status from
pending-smoke-telemetry to the resulting final evidence status.

In `@scripts/run_publication_matrix.py`:
- Around line 535-540: Keep the smoke-row state logic in
scripts/run_publication_matrix.py lines 535-540 based on current-run,
current-cap raw/checkpoint artifacts rather than manifest acceptance; expose
manifest acceptance separately or validate its artifact path/hash and cap before
using it. In tests/test_publication_runner.py lines 642-646, key the reservation
with cell.cap_label and assert the row-level reservation. In
tests/test_publication_runner.py lines 673-676, write raw artifacts under the
current 4,096-token stem so the monitor reads them.
- Around line 787-790: Update the reasoning-token validation around the
reasoning_policy checks to require reasoning_tokens to be a non-negative
integer, explicitly excluding booleans. Apply this validation to the
mandatory-minimum telemetry check and ensure negative values are rejected
wherever reasoning telemetry is validated, while preserving the existing
disabled-policy handling.
- Around line 995-999: Update the smoke reuse branch in the publication matrix
flow to skip reuse when args.preflight_only is enabled, ensuring that
_validate_openrouter_endpoint still runs for explicit live endpoint preflight
while preserving normal reuse behavior otherwise.
- Around line 87-100: Update the reasoning-policy validation around policy and
effort checks: require mandatory-minimum models to provide a non-null
reasoning_effort and matching OPENROUTER_REASONING_EFFORT, and reject disabled
models when fixed_options contains any reasoning-effort setting in addition to
enforcing disabled reasoning. Preserve the existing unknown-policy validation.

In `@web/scripts/build_leaderboard.py`:
- Around line 116-125: Update the upstream provider handling in the
expected_options construction to access registered["upstream_provider_slug"]
safely via the existing optional-value pattern, while preserving the current
OPENROUTER_PROVIDER_ONLY validation behavior when the slug is present.

In `@web/src/data/leaderboard.json`:
- Around line 83-86: Regenerate the leaderboard gate data from the frozen lane
and current model registry: update frozen_output_token_cap to reflect the
4,096-token policy and replace the obsolete embedded three-model list with the
current registry entries. Preserve the empty missing list and
model_registry_frozen state, and apply the same regenerated values to the
additional leaderboard records noted in the comment.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@scripts/run_publication_matrix.py`:
- Around line 803-806: The mandatory-reasoning validation in
gm_bench.publication.smoke_manifest_issues must reject reasoning_tokens when it
is not an integer, is a boolean, or is negative, matching _record_smoke_issues
before allowing panel unlock. Update scripts/run_publication_matrix.py lines
803-806 accordingly; add coverage in tests/test_publication_runner.py lines
570-599 through the persisted-manifest or build_cells("panel") path for True and
-1.