Dedupe identical token analysis payloads on PT9 import

PT9 import writes its converted TextAnalysis straight to storage (convertPt9Project.ts:163-170 → savePt9Import), never reaching the find-or-create in appendApprovedAnalysis (src/store/analysisSlice.ts:253-266). Two tokens the import glosses identically therefore get two payloads, where a hand-glossed pair gets one shared payload at frequency 2.

Reproduction. PIA (test-data/pt9-projects/PIA, #283) approves four plovs tokens across PHP 1:6-1:7. The pool key plovs holds four entries, each frequency 1:

The first two are analysesAreIdentical and render as indistinguishable rows in the suggestion dropdown.

Consequences.

• The frequency ranking in comparePoolEntries (suggestion-engine.ts:77-80) never sees a count above 1 on an imported project, so the id tiebreak decides every homograph.
• #186's panel is already on main (AnalysisCatalogPanel, selectCatalogRows), so this is live rather than anticipated: every PT9-imported project renders each twin as its own row at one usage, against #186's own "one row per distinct token analysis".
• Storage bloat, proportional to how repetitive the source is.

What to do

One content-dedupe pass in convertPt9Project.ts, just before TextAnalysis is assembled — both merged.tokenAnalyses and barePayloads exist by then, so one pass covers both. Inside mergeLanguageAnalyses would miss the bare payloads.

Take (tokenAnalyses, tokenAnalysisLinks), return them reduced:

• Bucket by normalizeSurfaceForm(surfaceText), then find within the bucket with analysesAreIdentical (both in utils/analysis-identity.ts). A flat linear find is O(n²) — fine for the store's one-write-at-a-time growth, not for a whole Bible. Bucketing is sound because analysesAreIdentical already requires equal folded surface; same structure as buildPoolIndex (suggestion-engine.ts:95-113).
• Keep the first payload, drop the rest, repoint every link's analysisId — what mergeIntoIdenticalPayload (analysisSlice.ts:353) does after an edit.
• Leave the links otherwise untouched: token.surfaceText stays each token's own, which is what keeps per-token drift detection accurate.

Run this pass after #324, which strips blank morpheme forms at import — stripping a morpheme turns previously-distinct payloads into twins, so the content fold must see the final shape.

Add identicalPayloadsMerged to Pt9MergeReport (report.ts:54) and seed it at report.ts:159. No UI change — Pt9ImportModal reads only report.languages.

Why the fold stays

Considered and rejected: an import-only exact-surfaceText rule instead of analysesAreIdentical's case fold.

Against it — appendApprovedAnalysis folds, so a later hand-gloss would adopt whichever duplicate sits earlier in tokenAnalyses, splitting the links on array order. buildPoolIndex folds too, so the sentence-initial twin would stay an indistinguishable extra row. And sentence-initial is the common case in scripture, so the rule would decline to merge precisely where it matters most.

Its one real cost, accepted: the surviving payload's surfaceText is whichever occurrence came first. That is presentation, not identity, and hand-glossed data already has it — appendApprovedAnalysis's own doc describes "a sentence-initial form links to a payload first created from a mid-sentence form". The fix belongs in #186: derive a row's displayed surface from its links — the most common token.surfaceText among them — recomputed on read rather than stored, since any stored choice goes stale on the next write.

Checked, no work needed. The generated ids (pt9:ta:, pt9:wa:, pt9:pa:) and producer are write-only, so an adopted payload keeping the first token's id is safe. clusterParseIdentities is a Set keyed by folded surface plus signature and already collapses duplicates. approvedTokenSeen runs before payload creation, so the one-approved-per-token invariant is unaffected.

Decide.

• Keep buildBareWordAnalyses' own seenByWord check (bareWordAnalyses.ts:52-64)? It keys on the raw word and the lexeme-id signature, so a content pass is strictly wider. Keeping both leaves skippedExistingIdentical its current meaning.

Tests. Twins collapse with both links surviving; a bare payload identical to a linked one; case variants collapse; and the three near-misses that must not collapse — different gloss, breakdown vs none, different breakdown. PIA is the manual counterpart.

Related: #187 (same gap for phrases, on the store write path), #186 and #254 (both compute against pool-bucket contents and assume payloads are deduped).

Size: S

Start in convertPt9Project.ts after #324's import changes, then read analysesAreIdentical and normalizeSurfaceForm in utils/analysis-identity.ts, plus Pt9MergeReport in report.ts. Verify the PT9 import tests cover linked and bare twins, case variants, and the three near-misses; done means duplicates collapse with links preserved and identicalPayloadsMerged is reported without UI changes.