finalize: rename --qat-convert to --quantize (unified QAT + PTQ)

[jdinalt]

Summary

• Rename forgather finalize --qat-convert <recipe> → --quantize <recipe>. The convert pipeline already re-installs fake quantizers on the loaded float weights, so it works on any source — the old name no longer matched the behaviour.
• The same flag now documents two modes: QAT round-trip (source trained with --qat-recipe, keeps the QAT accuracy benefit) and PTQ (plain bf16 source, standard post-training quantization).
• Promotes the previous "Behavior on Models Without QAT" footnote in docs/trainers/qat-training.md to a first-class PTQ Mode section with its own example.
• No backwards-compat shim for the old flag.

Closes #40.

Why this is a rename, not a new flag

PR #39 verification surfaced that Forgather's sharded checkpoint saver returns plain float weights from FakeQuantizedLinear (scale/zp inner state is non-persistent). The fix in #39 made _apply_qat_convert always run prepare → convert on the loaded weights — which incidentally also produces a valid PTQ artifact when the source was plain bf16. Issue #40 reframed from "enable PTQ" to "fix the flag name". Single flag, two modes, documented as such.

Marker-file infrastructure (proposed in #41 for load-time autodetection) is deferred — not needed for this rename and would couple #40 to the eval/inference issues.

Rename surface

Layer	Old	New
CLI flag	--qat-convert	--quantize
argparse dest	qat_convert	quantize_recipe
job_params key	qat_convert	quantize
Server passthrough param	qat_convert	quantize
finalize_model.py fn	_apply_qat_convert	_apply_quantize
TSX state / constant / label	qatConvert / QAT_CONVERT_RECIPES / "QAT Convert"	quantize / QUANTIZE_RECIPES / "Quantize"

Test plan

• forgather finalize --help shows --quantize; --qat-convert is gone.
• grep -rn "qat[-_]convert\|qatConvert\|QAT_CONVERT" src/ tools/ templatelib/ docs/ returns no hits.
• Dry-run smoke (--dry-run --quantize int8-dynamic-act-int4-weight) on a real model prints "Would run quantize step with recipe 'int8-dynamic-act-int4-weight'".
• PTQ end-to-end on the bf16 chinchilla baseline: 29 IntxUnpackedToInt8Tensor weights on disk (the new documented PTQ path).
• QAT round-trip path is structurally unchanged — the only diff is the user-facing flag name. (Already verified in torchao QAT support: --qat-recipe (trainer) + --qat-convert (finalize) #39.)
• ./build-webui.sh succeeds; FinalizeModal dropdown labeled "Quantize".
• Webui submit through the rebuilt FinalizeModal lands a job whose argv includes --quantize <recipe> (user-side check).

🤖 Generated with Claude Code

[jdinalt]

Review pass complete

Two review agents ran in parallel. Both said no blockers.

Applied in commit ca80b4e

• Dropped the stale `--safetensors` parenthetical from the Quick Start (the example no longer mentions `--safetensors`, so the note was orphaned). Replaced with a direct forward-reference to the Save Format section.
• FinalizeModal dropdown placeholder: `(none — skip quantize)` → `(none — no quantization)` (the verb form scanned awkwardly).

Filed as follow-ups

• finalize: warn when QAT-trained source is finalized without --quantize #44 — finalize should warn when run on a `--qat-recipe`-trained source without `--quantize`. Today the saver strips fake-quant inner state, so the default finalize silently throws away the QAT training-time benefit. This is the marker-file mechanism originally proposed in forgather eval: load and evaluate torchao-quantized models #41 / inference server: load torchao-quantized models #42, applied to the finalize side.
• Quantization: namespace recipe strings for future non-torchao backends #45 — namespace recipe strings (`torchao:int4-weight-only` or `--quantize-backend torchao`) before users script against bare names. Cheap to lock in now while the flag is new.

Skipped with reasoning

• TSX localStorage migration of the old `qatConvert` key — reviewer Add comprehensive unit tests for forgather modules #1 explicitly said "not worth holding the PR" and the field falls back to `""` cleanly.
• PTQ accuracy-claim hedge — one reviewer flagged, the other approved the existing hedging ("how much depends on the model and recipe"). Calling tie in favour of the existing wording.
• `QAT_RECIPES` → `QUANTIZE_RECIPES` Python rename — the module docstring at `src/forgather/ml/qat_recipes.py` already clarifies both modes. Reviewer Add claude GitHub actions 1771648604662 #2 said the docstring is cheaper than the rename.
• `--quantize` table-cell length in `finalize-model.md` — reviewer Add claude GitHub actions 1771648604662 #2 said "not blocking; content is correct".