Analysis Catalog: morpheme rows

Add a morpheme list to the Analysis Catalog (#186), so morphemes can be counted, sorted, and edited independently of the words that contain them.

What is already reachable without this

Morpheme data is not invisible in the token-only catalog: morphemes render as detail inside a token row, and token search matches morpheme forms and glosses. Typing DAT.SG already finds the words that use it.

What is missing is everything that needs a morpheme to be its own row:

• independent counts and sorting ("how many times have I used DAT.SG across all words?")
• filtering to bare-form morphemes — split out but never glossed, which is real unfinished-work triage
• a global morpheme gloss edit that fixes every containing analysis at once

Why this is the hardest of the three lists

MorphemeAnalysis is not a payload. It is a nested array inside TokenAnalysis.morphemes, with an id unique only within that array, no link table, and no dedupe. So:

• Identity has to be derived. Group by the projection that already exists — morphemeIdentity in analysis-identity.ts (form + gloss + entryRef / senseRef / allomorphRef / grammarRef). This makes the catalog its second consumer; keep the two in step.
• Usage locations are a union. A morpheme row's locations are the locations of every containing token analysis.
• Editing fans out and can cascade. Writing a morpheme gloss into M containing TokenAnalysis payloads can make several of them content-identical to other analyses, firing mergeIntoIdenticalPayload. Editing one morpheme row can silently collapse several token rows into each other. That has to be surfaced, not silently applied.

Delete semantics: clears the analysis, keeps the parse

Deleting a morpheme row removes the analysis, not the morpheme. The M morphemes lose their gloss and lexicon refs; form and position stay. The row -ῇ -> DAT.SG (40) disappears and -ῇ -> (unglossed) (40) appears. Every parse stays intact.

Rejected: removing the morpheme from each parse. That leaves M breakdowns that no longer decompose their word, and it fires the cascade above as a second surprise from one click. Re-cutting a word is what the morpheme editor on the token chip is for, where the whole word is visible.

This keeps delete meaning one thing everywhere: it removes recorded analysis, never the shape of the text or its decomposition.

Note on breakdown editing

Editing how a word is cut stays on the token row, not here — that is a property of the token analysis, and one edit there already fixes every usage. A morpheme row edits a morpheme's gloss and refs. Two different questions about two different objects, which is why they are separate lists.

Follow-up to #186.

Start with analysis-identity.ts and the existing Analysis Catalog work from #186, then trace how TokenAnalysis.morphemes and mergeIntoIdenticalPayload are handled. Define the morpheme-row identity and usage union, and ensure independent counts, sorting, filtering, gloss/ref editing, and analysis-only deletion are represented. Done includes surfacing cascades when edits merge token analyses.