[FR-011] Maintainer can make resume HASH=<x> after orchestrator crash

[explosivebit]

Trace: PRD prd-v0-1-smoke-evaluation-run · FR-011 · NOTE-001 crash recovery strategy · RFC AC-6 · RR-6

Capability: Load manifest.journal.ndjson, diff vs expected grid (model × task × seed combinations from run_hash inputs), schedule missing eval_ids only.

Acceptance:

• AC-6 from RFC: journal loaded, missing eval_ids identified, only those re-scheduled
• Crash modes covered (per NOTE-001 table): mid-eval, between-eval-completion-and-journal-write, mid-aggregating, mid-rename, post-published
• Append-only journal — entries never modified

Implementation locus: apps/eval-core-py/src/orchestrator/{journal,resume}.py + Makefile resume target