convert: default to group-scaled int4 (gs64) — the #455 tracking item

[ZacharyZcR]

First brick of the 1.2.0 "serve path from working to reliable" arc: stop handing new users the per-row quality trap by default.

Why now

Per-row int4 is the most-reproduced failure source in the tracker: −9.3pp mean acc_norm in the ablation (#453), the root cause of the think-mode loops and EOS starvation in #455 (gs64 cured 5/5 capture-matrix cells), and the very first thing the newest field report (#549) tripped over — that user's first conversion was per-row, their "quality worse than Qwen" was exactly the measured per-row deficit, and #455 closed with "make the converter default to group scales" as tracking item (a). This PR is that item.

Changes

1. convert_fp8_to_int4.py: --group-size default 0 → 64. gs64 is the community-validated point (the GLM-5.2 int4 think mode: stochastic reasoning loops / EOS starvation on stock v1.0.0; #307's TEMP=0.9 NUCLEUS=0.95 stopgap makes it severe #455-curing container, the tools: IQ3_XXS-codebook scheme in the quant ablation — settles #452's codebook decision #453 ablation arm at −2.2..−3.4pp vs fp16). Per-row stays available as an explicit --group-size 0 opt-out.
2. coli convert: --group-size passthrough (default 64) so the launcher path is controllable without dropping to the tool.
3. docs/quickstart.md: the "Easiest" recommendation now points at the gs64 container (mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp) instead of the per-row one it recommended until now — the docs were steering every new user into the exact container class that GLM-5.2 int4 think mode: stochastic reasoning loops / EOS starvation on stock v1.0.0; #307's TEMP=0.9 NUCLEUS=0.95 stopgap makes it severe #455 documented as broken. The per-row cards stay mentioned, labeled with what they cost.

Safety

• No silent mixing on resume: check_or_record_params already records group_size in the outdir manifest, so a pre-default per-row conversion resumed under the new default refuses loudly with the existing "use a fresh --outdir" message (convert_fp8_to_int4.py: --mtp is silently ignored with --indir and overwrites the finished container with a full int8 conversion #355-class guard doing its job) rather than interleaving formats.
• MTP head unaffected: the int8 head path (bits=8) never enters the grouped-int4 branch, and coli convert's step-2 re-invocation inherits the flag harmlessly.
• Engine side is long since ready: fmt=4/gs64 is the format the tracker's production containers already use (cuda+engine: full fmt=4 (grouped int4 gs=64) support + diagnostic harness #298/fmt=4 CUDA attention crash — FIXED (#298); benchmark tracker + follow-ups (PROFILE accounting, cache warmth) #334/cuda: grouped-int4 (fmt=4) support in the expert-group kernels — opens the GPU tier to g64 and E8 containers (#334) #451 line, all merged); the issue-MTP: int4 head unusable (int8 fixes acceptance); speculation still a net loss on CPU streaming-MoE — measured. #8 eh_proj warning text updated to match the new default.