Fix run-integrity issues #25/#26/#27 from the Fable 5 live run

[jkbennitt]

Closes #25, closes #26, closes #27. All three were found via the Fable 5 live run's artifacts (results/fable5-live-N1).

#25 — threat_response unwinnable

• first_draft_tick is now actually written: successful draft executions record per-threat response delay in loop ticks (_record_draft_response); threats appearing while colonists are already drafted record an instant (0-tick) response.
• Null incident placeholders (enemy_count=0 && threat_level=0.0) no longer enter threats_seen — the live run's "threat" was a null /incidents entry, and DefenseCommander's refusal to draft was correct.
• SCORING_VERSION 1.0 → 1.1: v1.0 scores with non-empty threats_seen are not comparable (the Fable run's 0.754 ex-artifact recomputes to ~0.834).

#26 — wandering shelter

First successful get_terrain_summary is pinned for the client lifetime. Terrain is static; re-deriving the colony center from live pawn positions made the verified sites chase the builders (10 blueprints, 10 sites, 0 shelters, mood 0.52→0.36).

#27 — pawn-targeted writes 1/13

• work_priority accepts the three shapes models emit ({"<WorkType>": n} documented, {"work_type": X, "priority": n}, {"work_priorities": {...}}) — previously the second shape posted literal work="work_type", priority="Research" to RIMAPI, which is what the "null-ref" actually was.
• job_assign accepts the job alias and maps x/z to target_position; an absent job name now fails visibly (ActionOutcome error) instead of posting an empty JobDef.

Observability

ActionOutcome + ACTION_EXEC events now carry action parameters (root-causing #27 required exhuming payloads from raw LLM text).

Verification

• 440 tests pass (17 new across executor normalization, threat tracking, terrain pinning), ruff clean, mypy strict clean.
• Live 2-tick preflight pending RimWorld being loaded — will report results on this PR before merge.

🤖 Generated with Claude Code

[jkbennitt]

Live 2-tick preflight complete (results/fable5-preflight5, Fable 5, seed 42). All five verification targets pass:

Target	Before (live-N1)	After
threat_response with no hostiles	0.0 (phantom null incident)	1.000
Pawn-targeted write success	1/13	3/3 (both work_type and nested work_priorities shapes normalized; bed_rest also landing)
Shelter site across ticks	new site every tick	(133,138) both ticks — pin holds
ACTION_EXEC parameters	absent	present — and immediately useful: both remaining failures are fully diagnosable from the event alone
scoring_version	1.0	1.1

Composite 0.859 over 2 ticks, 14/14 deliberations parsed, zero retries, $0 on subscription.

Two new visible (non-silent) failures surfaced by the params observability, neither in scope here: stockpile_zone priority sent as the string "Important", and a t1 growing_zone rejected with "Invalid plant definition: Plant_Rice" while t0's 'plant': 'rice' succeeded — RIMAPI def-name quirk worth a small follow-up issue.