fix: unblock Crashlanded research scoring (seed bench + brief observability)

[jkbennitt]

Why

Scoring 1.2 research is completed / (completed + available). On Crashlanded that starts at 7/31 ≈ 0.2258 — the finished/available ratio already in rle_crashlanded_v1, not model variance. The seed had no ResearchBench and currentProj unset. Agents never finish tech; gates that force research_target get RIMAPI “bench/prereqs missing.” Do not treat that floor as σ.

CoS greenlit this exact scope: seed + brief observability + tests. No scoring formula / version bump, no baseline-delta, no matrix or billed cal runs.

What landed

Seed

• Built, player-owned wood SimpleResearchBench (Building_ResearchBench) at (128, 0, 136) in docker/saves/rle_crashlanded_v1.rws and the gzip sibling saves/rle_crashlanded_v1.rws.gz (round-trip gzip; BOM + CRLF preserved).
• Simple bench does not need a generator (CompPowerTrader is hi-tech only). Encoded like other built furniture in this save (stuff, faction Faction_11, empty billStack).
• Queued currentProj=Smithing using the same researchManager field rle_ship_launch_v1 already uses.
• Re-pinned save_sha256 on 01_crashlanded_survival.yaml. The .baseline.json sidecar is not recalibrated (scoring 1.2 / no baseline-delta).

Brief

• research_bench_present on the harness-neutral state snapshot.
• MAP_SUMMARY cue when the flag is false.
• get_map prefers research benches in the 50-structure sample so 400+ ruin walls cannot hide the seeded bench.

Executor

• research_target on the already-queued current project is success-by-state (no rewrite), same pattern as a covered growing zone.
• Scripted MCP smoke now targets Smithing (available, not current) so the write path is still exercised.

Tests

• 7/31 floor math
• brief flag true/false + missing-bench cue
• save contract (SimpleResearchBench + gzip == canonical .rws)
• current-project already-satisfied path
• get_map keeps the bench when there are 60+ walls

Out of scope

• Scoring 1.2 formula / SCORING_VERSION
• Baseline metric recalibration
• Rebuilding the five derived scenario saves
• Live / billed model runs

  

[jkbennitt]

LGTM — scope matches CoS greenlight: Crashlanded seed SimpleResearchBench + currentProj=Smithing, brief research_bench_present + MAP_SUMMARY cue, get_map prefers benches, research_target current-project success-by-state, smoke targets Smithing for write path. No scoring/version/baseline bump. Tests cover floor math, save contract, brief flag, already-satisfied, get_map keep. All five CI SUCCESS on 7c61cd8 (run 34237372223). HOLD merge for CoS clear. Same Main 1-review gate as #74 (own-PR APPROVE impossible as jkbennitt).