fix(benchmark): bind CALF/LegNav paired slots to the observed observation contract

[ll7]

Goal / Problem

The CALF/LegNav comparator added in #7357 assigns each paired trace to a fixed slot
(perfect_perception / sensor_limited) but never checks that the trace's own observed
observation contract matches the slot it was filed under.

robot_sf/benchmark/calf_legnav_comparator.py:

• _observation_contract() derives condition from the trace's observed_observation.evidence_class
  (ideal_state -> perfect_perception, perception_limited -> sensor_limited) and returns
  status: "available" for either.
• build_calf_legnav_comparator_report() then files perfect_trace under CONDITION_IDEAL and
  sensor_trace under CONDITION_SENSOR without comparing the derived condition to the slot.
• The paired identity guard covers candidate, scenario_id, seed — but not horizon.
• robot_sf/benchmark/schemas/calf_legnav_comparator_config.v1.json allows noise/occlusion keys in
  either condition, so a config edit can make both rows perception-limited.

Effect: a mislabelled or degenerate pair (for example both rows perception_limited) still reports
status: "available" at the top level. The mismatch is visible in the per-condition
observation_contract.condition field, but nothing fails closed. PR #7357's own body asserts "both
rows use the same candidate, scenario, seed, horizon, and config digest; only the observation
contract differs" — that invariant is currently asserted rather than enforced.

This is a fail-open in the exact property the PR is being domain-reviewed for. The committed smoke
config is correct today (verified: perfect_perception -> ideal_state / noise_profile: none;
sensor_limited -> perception_limited / bounded_gaussian); the gap is that an edit could break
the contrast silently.

Scope

• In scope: bind each condition slot to the observed evidence class and fail closed on a mismatch;
  add horizon to the paired identity check; cover both with tests.
• Out of scope: any change to the comparator's claim boundary, evidence tier, or the multi-seed
  follow-up in research: extend CALF/LegNav diagnostic to multi-seed sensor and embodiment study #7356.

Prepared implementation

A validated implementation is already pushed to branch
review/pr-7357-observation-binding-20260818, cherry-picked onto PR #7357's head
c680b6811b6b8d1b8f750ec9c1892c7d433d9901 (single commit 5949b7024).

It threads an expected_condition through _observation_contract() / _condition_report(), records
expected_condition and condition_binding (matched / unavailable / not_checked) plus a
reason in the report, sets the condition status to unavailable (and therefore the condition and
report status to blocked) on a mismatch, and adds horizon to identity_fields. The report schema
needs no change: condition.observation_contract is additionalProperties: true.

Validation on that branch:

• pytest tests/benchmark/test_calf_legnav_comparator.py tests/dev/test_issue_7331_benchmark_namespace_inventory.py — 12 passed
  (including three new tests: swapped contract fails closed, matched binding recorded, horizon mismatch rejected).
• ruff check / ruff format --check on the changed files — passed.
• Live paired smoke on configs/benchmarks/issue_7318_calf_legnav_comparator_smoke.yaml — still
  status: available, both slots condition_binding: matched.

The work was parked rather than pushed to research/issue-7318-calf because the autonomous factory
pushed a base refresh to that branch minutes earlier and was mid-review; racing its exact-head verdict
was avoided deliberately.

Definition of Done

• A paired report whose slot and observed evidence class disagree is blocked, not available.
• A paired report whose two rows differ in horizon is rejected.
• The committed smoke config still produces an available report with both slots matched.
• Either the branch above is merged into feat(research): add CALF/LegNav comparator diagnostic #7357 before it lands, or the change lands as a
  follow-up PR after feat(research): add CALF/LegNav comparator diagnostic #7357 merges.

Validation / Testing

• uv run pytest tests/benchmark/test_calf_legnav_comparator.py
• uv run python scripts/benchmark/run_calf_legnav_comparator_issue_7318.py --config configs/benchmarks/issue_7318_calf_legnav_comparator_smoke.yaml --output-dir <tmp>

Project Metadata

• Priority: P4 — small, bounded, and it closes a fail-open in a diagnostic contract.
• Effort (h): 0.5 (implementation exists and is validated; needs review + landing).

[ll7]

Reconciliation status

A validated implementation already exists at commit 5949b7024 on review/pr-7357-observation-binding-20260818. PR #7357 is the preferred landing surface. Do not start another implementation.

Required landing sequence

1. Refresh the prepared commit onto the current exact head of feat(research): add CALF/LegNav comparator diagnostic #7357, or apply it as the immediate successor if feat(research): add CALF/LegNav comparator diagnostic #7357 merges first.
2. Require a fresh exact-head review after the commit is present; prior review evidence does not cover the new binding checks.
3. Preserve the report vocabulary and evidence boundary. A slot mismatch must make the condition and top-level report unavailable/blocked; it must not be converted to a zero-valued result.
4. Add horizon to the paired identity contract together with candidate, scenario, seed, and config identity.
5. Re-run the committed paired smoke and verify both expected bindings are matched.

Dependencies

• Implementation parent: research: reproduce CALF/LegNav-style RL local-navigation evaluation #7318 / PR feat(research): add CALF/LegNav comparator diagnostic #7357.
• Multi-seed external-artifact follow-up: research: extend CALF/LegNav diagnostic to multi-seed sensor and embodiment study #7356; it remains separately blocked and is not completed by this fix.
• No schema expansion is required unless review finds that the current additionalProperties policy is insufficient.

Close condition

Close after the change is merged, the swapped-condition and horizon-mismatch tests pass at the merged exact head, and the canonical smoke remains available with both slots matched. This is diagnostic contract integrity, not CALF/LegNav transfer evidence.

[ll7]

Claiming the prepared landing path for this research-contract hardening issue.

• claim ref: agent-claims/issue-7454
• base: 8272c3af79b63a72827010d02801315c8c8cf152
• prepared implementation: 5949b7024fbddfc1e021bc49db5f87e0e2132be9 on
  review/pr-7357-observation-binding-20260818
• landing target: PR feat(research): add CALF/LegNav comparator diagnostic #7357 / research/issue-7318-calf
• scope: land the existing slot-binding, horizon-identity, and focused fail-closed tests after a
  current-main refresh; preserve the diagnostic-only boundary and pending domain approval
• non-scope: compute, training, campaign execution, benchmark ranking, transfer, safety, or paper
  claims

No duplicate implementation will be started; the claim is released if PR #7357 has another active
writer or if its maintainer/domain ruling changes.

[ll7]

State synchronization — 2026-08-19

PR #7357 is open and owns the prepared implementation at head ec2f61aca6db146345471537663a146d2d4202fb. The issue is therefore state:running, not state:ready.

The remaining gate is domain-aware approval plus exact-head review and hosted CI. Do not create a competing follow-up PR while #7357 remains open. If #7357 closes without merge, return this issue to ready only after revalidating the prepared commit against current main.