Merge pull request #20 from AppSprout-dev/feat/restore-deliberations-export

Phase A observability floor: deliberation logs, replay metadata, timeouts, cost source

Restores per-commit history for #20 (originally landed as a single squash commit
61bce01; this merge re-establishes the 9 individual commits + the cross-platform
.gitattributes CI fix as separate parents-of-master so they remain bisectable).

* Restore per-scenario *_deliberations.jsonl export in run_scenario (A1)
* Expand DELIBERATION event payload with actions + summary (A2)
* Surface raw LLM output via PROVIDER_CALL events (A3)
* Add replay-grade metadata + --seed flag + per-scenario summary JSON (A6.1)
* Pin canonical save SHA-256 into scenario YAMLs + loader validation (A6.2)
* Per-task LLM timeout via asyncio.wait_for (A7)
* Populate benchmark_history.jsonl from run_scenario.py too (A8)
* Cost tracker: surface pricing source + manual override flags (A9)
* Treat .rws saves as binary so scenario SHA pins are cross-platform (CI fix)

Author: jkbennitt

.gitattributes

@@ -0,0 +1,6 @@
+# RimWorld save files are XML text but we treat them as binary blobs so the
+# canonical save_sha256 pinned in each scenario YAML matches across platforms.
+# Without this, Git's autocrlf=true on Windows converts CRLF<->LF on
+# checkout/commit and the on-disk byte sequence (and thus its SHA-256)
+# differs between OSes, breaking the scenario loader's save validation in CI.
+*.rws binary