De-voice writer SFT pairs and widen the ship-gate holdout

[telivity-otaip]

Summary

The writer SFT built each row's brief by extracting sentences from the post it
was trying to teach the model to write. Measured on the corpus, the share of
brief words sitting inside a 5-gram of their own target had a median of 1.0,
and 126 of 129 rows would themselves have tripped the brief_echo_reject limit
the gate uses to disqualify a draft. Copying the prompt forward was a winning
strategy for that objective.

This PR makes the pairs (D(y), y) and gates them on that property.

De-voicing operator (devoice.py)

Deterministic and inspectable — no model in the loop, since an LLM flattener is
exactly the component that can leak cadence back into the input.

• kept: named entities, evidence figures, claim vocabulary, clause order
• removed: second-person address, contractions, emphasis punctuation,
  shouted words, sentence-initial conjunctions, discourse markers, one-line
  fragment rhythm, and the article/copula/preposition scaffolding that separates
  a written sentence from a jotted note

Negation and comparison words are deliberately never dropped: removing not
would not de-voice a claim, it would reverse it.

The operator ships with its verifier. Each pair is measured on the existing
gate_pair cadence axes, checked for entities or figures the operator itself
introduced, and dropped unless the brief's overlap with the post stays under a
cap. Fail-closed — a row that cannot be moved is not trained on.

brief→post 5-gram copy ratio	median	p90	max
before	1.0	1.0	1.0
after	0.147	0.263	0.333

build-writer-sft now reports the ratio distribution and a per-reason drop
breakdown instead of one opaque skip count.

Widened holdout carve (select-writer-holdouts)

The previous gate ran on three holdouts. Three paired comparisons cannot reach
any significance threshold — a clean sweep is still p=0.125 — so the result
could not distinguish a regression from a coin flip.

The carve reserves a share of the briefable pool (25%, clamped to 12–24),
orders candidates by a stable digest of the piece id so it is reproducible, and
keeps previously carved ids pinned so runs stay comparable. Pieces that cannot
produce a brief are excluded: neither arm could be scored on them.

index-voice --from-carve reads the ids from the carve file, since retyping
twenty flags is how a holdout quietly re-enters retrieval.

Committed ship gate (eval-writer-adapter)

Both arms see the same de-voiced brief and the same exemplars; only the adapter
differs. An adapter is kept only when all three hold:

1. it wins more holdouts than it loses
2. the margin clears a one-sided sign test
3. it is not disqualified more often than the arm it replaces

Weights are loaded once per arm rather than per draft. --archive-on-fail moves
a rejected adapter aside so the write path resolves back to adapter=none.

Also

• train --detach runs a train in its own session via start_new_session.
  Shell-level setsid is util-linux and absent on macOS, where a launcher that
  reaches for it fails before Python starts.
• Writer train recipe separated from the translator defaults: 16 layers, LoRA
  rank 16, lr 3e-5, 10 epochs. Rank and learning rate are plumbed through the
  chunk worker and recorded in the checkpoint meta.
• release_mlx_memory now honours the MLX opt-in gate. A Metal-less session
  aborts inside metal::load_device via C++ terminate, which the surrounding
  except Exception cannot catch.
• mine_brief_from_holdout derives its minimum source length from the caller's
  overlap budget instead of a fixed constant.

Test plan

• pytest — 317 passing, Contoso fixtures only
• ruff check src/ tests/
• build-writer-sft on a real corpus: pair copy ratio median 1.0 → 0.147
• select-writer-holdouts + index-voice --from-carve: holdouts confirmed
  absent from retrieval
• train --writer on the de-voiced SFT, then eval-writer-adapter on the
  widened holdout — adapter kept only if it clears the bar

No corpus, posts, or adapters are committed; profile data stays under
~/.personality-protect/.

Made with Cursor

[telivity-otaip]

Ran the full pass end to end on a real corpus.

Pairs — 60 gated (D(y), y) rows from 274 posts. Brief→post 5-gram copy
ratio: median 0.147, p90 0.263, max 0.333 (was median 1.0). Drops by reason:
137 under the 50-word floor, 29 over the copy cap, 20 holdouts, 10 fragment-gap,
9 operator-introduced entity/figure, 6 unbriefable, 3 sentence-length gap.

Train — 600 steps in 12 chunks at 16 layers / rank 16 / lr 3e-5. Peak Metal
34.1 GB against a 16 GB wired cap; train loss 2.08 → ~0.08.

Gate (n=20, adapter vs RAG-alone)

	adapter	rag	tie
wins	4	11	5
disqualified	12	6

p = 0.98 one-sided, so all three keep-conditions fail. The adapter was archived
and the write path resolves to adapter=none.

Worth recording from the disqualification breakdown, since it is what the next
pass should act on:

• brief-echo disqualifications went to 0 for both arms. That failure mode
  came from briefs being extracts of their targets, and it is gone.
• invention rose 5 → 11 and parroting 2 → 5 for the adapter. Train loss was
  already ~0.09 by roughly 3.5 epochs on 60 rows, so 10 epochs looks like too
  many for a set this size.

That hypothesis could not be tested from this run: the chunk trainer passes
--save-every per chunk, so every chunk writes 0000050_adapters.safetensors
and overwrites the previous one. Retaining per-chunk checkpoints is a
prerequisite for tuning epochs without a full retrain each time.

The gate result does not change the case for this PR — the pair construction,
carve and gate are the parts that make the next attempt measurable. It does mean
no adapter ships from it.

[telivity-otaip]

Rebased onto main (merged #49). Added durable per-chunk checkpoints, writer recipe at 3 epochs, and --sweep-checkpoints. Adapter weights still do not ship — write stays adapter=none until eval-writer-adapter returns keep.