research: extend CALF/LegNav diagnostic to multi-seed sensor and embodiment study

[ll7]

Question

Can the paired Robot SF observation-contract diagnostic from #7318 support a reproducible multi-seed, multi-scenario study of sensor-limited versus ideal policy behavior without being mistaken for CALF/LegNav transfer evidence?

Why this is separate from #7318

The #7318 fixture smoke validates trace separation, fixed-shape policy inputs, provenance, and local metric plumbing only. It is one seed and one scenario, and it does not hydrate the external CALF checkpoint, LegNav simulator, leg-sensing calibration, gait model, or TurtleBot 4 runtime.

Acceptance criteria

• Freeze a versioned scenario manifest with at least three seeds and a held-out scenario split.
• Keep policy, observation contract, simulator assumptions, and embodiment assumptions separate in every row.
• Report direct local metrics separately from qualified simulator-distance proxies, with uncertainty and sample counts.
• Exclude fallback/degraded/unavailable rows from any success-like aggregate and fail closed when the external source artifacts are unavailable.
• State Social Force/HSFM assumptions explicitly; do not infer sensor realism, safety, planner ranking, or zero-shot transfer.
• Produce a decision note that classifies the result as continue, revise, or blocked and names the next synthesis surface.

Stop rule

Stop as blocked before a broader run if the real manifest, source-policy artifacts, or required sensor/embodiment assumptions cannot be sourced and checksummed. Do not substitute the one-seed fixture for the missing proof.

Related work

• Parent research issue: research: reproduce CALF/LegNav-style RL local-navigation evaluation #7318
• Current implementation PR: feat(research): add CALF/LegNav comparator diagnostic #7357

schema: goal_autopilot_preparation.v1
repository: ll7/robot_sf_ll7
issue: 7356
source_body_sha256: 59bcf080399dc93d3d69f528d47a264e241ae1603847b5635aba068d5d697e63
source_comments_sha256: 
audit_schema: open_issue_contract_audit.v1
audit_digest: 1c191411793929640684e0827d763f3a2009a8682ab98dcefa8c2ded3e54f4ea
audit_classification: state_conflict
next_action: reconcile_state_labels
authority: lifecycle_owner
execution_mode: lifecycle
preferred_worker: MaxRunner
expected_pr_runner_label: runner:max
implementation_admitted: False
state_ready_change_proposed: False
mutation_batch: open-issues-20260830-03

This packet is preparation evidence only. It never overrides live labels, exact claim state, branch state, typed dependencies, domain gates, compute authority, release authority, or scientific evidence rules.

[ll7]

Revival contract — 2026-08-19

This multi-seed study is blocked on external source and embodiment inputs. PR #7357 owns the bounded one-seed observation-contract implementation; it does not provide the CALF checkpoint, LegNav simulator identity, leg-sensing calibration, gait/embodiment model, TurtleBot runtime, or a checksum-covered held-out manifest.

Unblock only when those inputs have exact source revisions, licences, hashes, observation/embodiment assumptions, and a versioned scenario/seed split. Then create a campaign-design child that freezes direct versus proxy metrics, exclusions, uncertainty, resource budget, and an explicit non-transfer claim boundary before any execution. Do not run the multi-seed study from this issue or infer CALF/LegNav transfer from Robot SF rows.