fix(catalog): preserve recurring outcomes as latest

[Kaileshwar16]

Fixes a catalog state bug where an outcome that occurred earlier could be deduplicated incorrectly when it happened again later.

For example:

error → success → error

The final error was previously treated as an old duplicate, so the catalog stayed ok and the episode remained in the dataset.

The same problem happened in reverse when a successful result followed an earlier failure.

This change keeps consecutive identical retries idempotent, while recording a recurring outcome as a new catalog occurrence. It also adds an optional execution_id for retries that may arrive after other executions have already completed.

The Parquet format and canonical episode identities remain unchanged.

Tests

Added coverage for:

• recurring outcomes
• consecutive retries
• selective-stage runs
• explicit execution IDs
• local and bucket-backed catalogs
• concurrent appends
• crash recovery and dependent-table repair
• historical timestamp preservation

Validation:

2032 passed, 9 skipped
ruff check --fix
ruff format
ty check
git diff --check

Closes #540.

[kstonekuan]

One gap: removing the execution_id must be non-empty when supplied check passes all 150 tests.

Everything else is held well. Reverting the scoping fails 21, and ignoring execution_id fails its own case.

I checked the property the new fingerprint could have broken, since it now depends on catalog state at append time: CATALOG.md:114 promises concurrent duplicate appends converge. Two workers appending the same outcome at once still produce one row, and so does a sequential repeat. Your design holds it.

[Kaileshwar16]

Thanks for the review. I removed the execution_id validation as suggested and reran the tests. The 150 catalog, dataset, and bucket tests passed, along with ruff, ty, and git diff --check.
I have pushed the follow up commit.

[kstonekuan]

I meant add a test, not delete the guard. My wording was ambiguous.

Without it, empty and whitespace ids become distinct scopes rather than an error:

None  -> out
""    -> out.a8de770a92e6a396
"  "  -> out.c758dd61b918da88

So execution_id="" silently buys a unique execution identity, and "" and " " are different ones. Restore the check and give it a case.

[Kaileshwar16]

@kstonekuan Sorry, I misunderstood the first comment. Restored the guard and added tests for empty and whitespace-only IDs.

[kstonekuan]

LGTM, merging.

Guard restored and held, including whitespace-only: dropping .strip() fails its case on its own.

No apology needed, my first comment was ambiguous.

Full suite on the rebased result: 2054 passed / 8 skipped.