Fix Edison sidecar provenance, drop malformed citation sidecar, KNOWN_BLOCKED provider

[realmarcin]

Summary

Three fixes, closing #641 and addressing part of #658:

1. scripts/_edison_capture.py: capture_full_response's sidecar_files provenance block used a plain .exists() sweep of the output directory. Edison output stems are deterministic, so a rerun of the same record/job lands in a directory already holding the previous run's sidecars — a rerun whose verbose fetch or list_files call fails wrote agent_state_json: true into a meta stamped with a new task_id, pointing at a trace belonging to the old one. Ported TraitMech's fix (via CultureMech#291's canonical, now-tested implementation): track which keys the invocation actually wrote, report the AND of "we wrote it" and "it is on disk". enrich_edison_response's disk-snapshot call site is unchanged, and a test now pins why it's allowed to differ.
2. scripts/research_community.py: dropped --separate-citations. That sidecar is a regex over the report prose, not structured provider output, and it's malformed — TraitMech's ORNL Clostridium-Desulfovibrio-Geobacter trophic model: curate causal graph #249 found 353 sidecars with 194 broken markdown-link tails, 2,770 stray trailing commas, and 332 of 353 duplicating a reference. The report's own References section is the trustworthy artifact.
3. scripts/deep_research_provider.py: brought to parity with CultureMech#285's canonical copy (KNOWN_BLOCKED provider table, --allow/--no-paid, the recommendable() helper, duplicate-alias guard) — same change landing in proteintraitsmech#514 and MediaIngredientMech#418. (An earlier revision of this description cited CultureMech#317, which is a separate, unrelated open PR — corrected per review.)

This PR does not fully close #658: its second acceptance criterion ("either this script is in the vendored-sync contract, or a documented decision explains why it deliberately is not") is still open. That's a fleet-wide sharing-model decision (CultureMech#287 explicitly deferred it until real copies existed somewhere — they now do, across 5 repos) and is out of scope for this PR; #658 stays open to track it.

Test plan

• 6 new regression tests for the Edison fix, ported from CultureMech#291 (no Edison client constructed, no credits spent — response is a stub, client=None exercises the skipped-fetch path)
• Mutation-tested: reverting only the sidecar_files source line turns test_stale_sidecars_are_not_attributed_to_the_new_task red
• Full uv run pytest tests/ — 2553 passed, 89 skipped, 8 deselected, 0 failed
• black --check/ruff check/mypy src/ all clean
• No provider was called; no research credits were spent

Closes #641

🤖 Generated with Claude Code

[realmarcin]

Round-2 fix pushed addressing the review's one medium finding: restored `test_main_rejects_unknown_provider_argument` and `test_main_rejects_unknown_focus_argument`, dropped in the round-1 test-file reorganization with no replacement. `main()`'s validation itself was untouched; only its test coverage had lapsed. Mutation-verified.

Full suite: 2553 passed, 89 skipped.

[realmarcin]

Addressed both findings by editing the PR description (no code change needed):

• Corrected the canonical-source reference from CultureMech#317 (wrong — separate, unrelated open PR) to CultureMech#285 (the actual merged source).
• Dropped the B009 claim — verified this repo's `_edison_capture.py` already used `result.content`, not `getattr`, both before and after this PR. That fix was real in MediaIngredientMech's port, not this one; the description had copy-pasted it.
• Changed "Closes Port CultureMech's KNOWN_BLOCKED provider check into deep_research_provider.py, and register the script for vendored-sync #658" to "addressing part of Port CultureMech's KNOWN_BLOCKED provider check into deep_research_provider.py, and register the script for vendored-sync #658" and explained why: Port CultureMech's KNOWN_BLOCKED provider check into deep_research_provider.py, and register the script for vendored-sync #658's second acceptance criterion (vendored-sync contract membership, or a documented decision not to) is a fleet-wide sharing-model question CultureMech#287 explicitly deferred, and is out of scope here. Port CultureMech's KNOWN_BLOCKED provider check into deep_research_provider.py, and register the script for vendored-sync #658 stays open.

[realmarcin]

Backported a real bug fix from CultureMech#317: provider_adjustments: null (explicit YAML null) used to skip validation entirely and crash downstream with AttributeError instead of a clean ValueError. Fixed to match the sibling capabilities check's unconditional isinstance validation. Full suite: 2554 passed.

[realmarcin]

Backported the same fix as CultureMech#317/proteintraitsmech#514's latest round: test_a_measured_dead_provider_is_not_recommended never set a credential at all (making the whole assertion vacuous), and even fixed, the None-branch escape hatch remained. Fixed both, matching the sibling repos' final state.

[realmarcin]

Round-N fix pushed: added `test_recommendable_no_paid_actually_excludes_a_high_cost_row`. This fix already landed in CultureMech#317/MediaIngredientMech#418/proteintraitsmech#514 earlier in the same porting effort — it was missed here specifically. Mutation-verified (deleting the `no_paid` branch from `recommendable()` leaves the full suite green without it).

Full `uv run pytest tests/` — 2555 passed, 89 skipped.

[realmarcin]

Round-N fix pushed: added `test_recommendable_allow_actually_excludes_a_disallowed_row`, same gap as the `--no-paid` one fixed last round (already landed in the sibling repos but missed here). Mutation-verified.

On the low finding (commit `b302bf3`'s claimed B009 fix doesn't correspond to any code change in this repo): confirmed, not amending that pushed commit message. This repo's copy never had the B009 finding to begin with — the claim was carried over from CultureMech's context where it did apply.

Full `uv run pytest tests/` — 2556 passed, 89 skipped.