Fix Atlas scans in text identity and cancellation recovery (JVNAUTOSCI-2727)

[witbrock]

Merge decision: ready — bounded live execution plans eliminate full-language text identity scans and empty cancellation-queue collection scans, with existing behaviour preserved by 106 targeted tests.

User outcome

Reduce recurring Atlas database work in text creation/reuse and interrupted-cancellation recovery. An English-NZ fingerprint miss previously examined 183,644 documents; an empty cancellation recovery poll examined all 1,074 queue documents.

Material changes

• Route fingerprint lookups, duplicate-insert reconciliation and text-relation removal through a repository lookup that explicitly implies the existing partial compound index's string predicate.
• Add a non-unique partial index on status, cancellation request time and ID, retaining the existing recovery filter and ordering.
• Preserve normalised/exact text identity, language matching and all cancellation decisions.

JVNAUTOSCI-2727 owns delivery status, including the separately attributed obsolete Windows embedding client. The corrected materialised embedding queue query already exists on main.

Evidence

• 106 targeted tests passed across text fingerprint identity, newline preservation, language-specific relation removal, duplicate recovery, queue behaviour, collection index setup and existing embedding queue semantics.
• Live read-only executionStats: English-NZ fingerprint miss 183,644 documents / 342 ms → zero documents / 0 ms; English miss 2,251 → zero; both select fingerprint_lang_unique.
• Canonical additive index setup materialised only status_cancellation_requested_id, preserving every existing index. The unchanged recovery query moved from COLLSCAN plus SORT, 1,074 examined documents, to the new index with zero examined documents and no blocking sort.
• Current embedding query uses embedding_status_updated_at_desc, examining one key/document for one result.
• Python 3.11 grammar check (1,440 files), PDM lock check and diff whitespace check passed. Black reports pre-existing formatting outside the changed lines in three files; no unrelated reformatting included.

Ship boundary

• Minimum ship criteria: preserve text identity and cancellation semantics while proving index-backed execution for the two changed paths; required CI passes.
• Stop-ship conditions: changed language/identity results, incorrect cancellation, destructive index replacement or no demonstrated scan reduction.
• Non-blocking observations: historical Atlas Query Shape Insights returns 403, so these cumulative counters do not establish the exact alert's cause. The old embedding query is from Windows and requires that client's update/restart; it is tracked in Jira and is not claimed fixed by this PR.
• Non-goals: alert suppression, query framework changes, semantic policy changes, unrelated search optimisation or broad remote deployment.