RUN_LENGTH=3 cannot see a retake that diverges at word three, which is six of the annotated defects

[Railly]

A speaker who restarts a phrase and changes its destination leaves a shared prefix and a divergence. The detector indexes candidates by a repeated run of three identical words (verify.ts:199, semantic.ts uses the same), so a retake that diverges at word 3 has no repeating trigram and is invisible.

Measured case

At ~91.5s of the Cueva recording:

"abrimos la guia para vincular, abrimos la opcion para vincular un dispositivo"

Normalised, the only runs occurring twice are the bigrams abrimos la and para vincular. No trigram repeats. Confirmed by running the detector over the real text: zero findings from findRepeatedPhrases.

Hunter's own annotated list of this material contains five more of this shape:

• "la ia tendria, la ia tendria acceso"
• "pero esta es muy interesante / pero esto es muy granular"
• "voy a agregar una feature para... para detectar"
• "para simplemente cifrar / para simplemente detectar con asteriscos"
• "voy a iterar normal para... para hacerlo mas cool"

Why this is not just a smaller RUN_LENGTH

Dropping to 2 words wholesale takes false positives from 8 to 64, nearly all of them 0-1 content-word connectives ("para que", "voy a", "de normal"). The content-word floor has to stay on the 2-word runs.

cli.ts:821 already describes this stacked-filler shape as intended behaviour; the implementation ships only the 3-word scan.

Also worth noting

growMatch (verify.ts:227) already computes the divergence point — it is the loop's exit condition — and discards it. For this class that index is the answer: the shared prefix runs up to it, and the second attempt's start is where the cut ends.

Depends on #71: with the intra-line skip in place, none of this is reachable anyway.

[Railly]

Partially addressed in 64d4bde (0.26.0). Leaving this open: the headline case is still uncovered.

What shipped

A new findRestarts in verify.ts, wired through mergeRepeats. It catches la ia tendria, la ia tendria acceso and para simplemente cifrar / para simplemente detectar, and refuses all three connectives this issue names plus the reuse #63 forbids.

Contrast run:

✗ catches a restart whose attempts diverge at word three
✗ catches the "la ia tendria" restart from the annotated list
 65 pass / 2 fail   → restored: 67 pass / 0 fail

The four negative-control tests pass in both states, so they are controls rather than vacuous.

Why the prescribed fix was not implemented

This issue asked for 2-word runs behind the content-word floor. Measured against the strings this issue itself names, that cannot separate the classes:

• real restart prefixes score 0, 1, 1, 1, 1, 2, 2 content words
• the false positives named here score 0, 1, 1

The distributions overlap completely. Scoring the grown span and the prefix-plus-both-destinations were also measured; neither separates (voy a mostrar / cerrar scores 3, above two real restarts). A structural rule shipped instead.

What is still open

abrimos la guia para vincular / abrimos la opcion para vincular — the abandoned attempt is 3 words against a 2-word shared prefix, outside RESTART_PROBE_LENGTH. Widening the bound needs a constant that could only be fitted to fixtures written for the purpose: the 417s corpus behind the 8-to-64 false-positive figure is not in this repo (checked examples/ and a repo-wide SRT search).

Reopening that needs the corpus, or a decision from the owner on the bound.

Related: #76, found while building a fixture here. y is missing from STOPWORDS_ES, which inflates every content-word score in the shared floor.