fix: size FP16 and ternary (TQ1_0/TQ2_0) GGUF quant types correctly

[SuperMarioYL]

What

_extract_quant_type() (the GGUF-filename → quant-type parser in models/fetcher.py) emitted tokens that the canonical QUANT_BYTES_PER_WEIGHT / QUANT_QUALITY_PENALTY tables don't key on, so two quant families were mishandled at fetch time.

1. FP16 GGUFs were sized ~3.5× too small

A *-FP16.gguf sibling extracts to the literal "FP16", but data/quantization.py keys full precision as F16 (there is no FP16 key). _estimate_gguf_size() does:

bpw = QUANT_BYTES_PER_WEIGHT.get(quant_type.upper(), 0.5625)  # default Q4_K_M

so the lookup misses and falls back to the Q4_K_M 0.5625 bytes/weight default. A 7B FP16 GGUF is then estimated at ~3.94 GB instead of 14 GB. Because the HF siblings listing carries no per-file size, _estimate_gguf_size() is the primary GGUF sizing path, so this affects essentially every FP16 GGUF — directly producing optimistic "fits" results, the opposite of what a hardware-fit advisor should do.

2. Ternary GGUFs were dropped entirely

*-TQ1_0.gguf / *-TQ2_0.gguf had no matching pattern in the extractor, so they returned "unknown" and were skipped by the if quant == "unknown": continue filter — even though the tables already fully price TQ1_0 and TQ2_0 (and list them in QUANT_PREFERENCE_ORDER). BitNet-class ternary builds therefore never appeared in results.

Fix

Additive, no table or heuristic changes:

• Add a TQ\d+_\d+ pattern to the extractor.
• Canonicalize full-precision aliases FP16 → F16 / FP32 → F32 to the keys the tables use (and add FP32 to the precision alternation so the alias is reachable).
• Regression tests, including a drift guard (test_extract_quant_type_keys_resolve_in_byte_table) asserting every quant the extractor surfaces from a real GGUF filename resolves in QUANT_BYTES_PER_WEIGHT, so the extractor and tables can't silently diverge again.

The quality-penalty path is unchanged: F16 already resolves to 0.0 in QUANT_QUALITY_PENALTY, so only the size estimate is corrected.

Testing

uv run pytest -q     # 400 passed
uv run ruff check    # clean

The four new tests are red on main and green here. No GPU required.

Notes

• Scope is limited to models/fetcher.py + tests; the separate non-GGUF (engine/quantization.py) FP16-keyed vocabulary used for repo-name inference is intentionally left untouched.
• "Does a real repo actually ship *-FP16.gguf / *-TQ1_0.gguf?" — yes; these are common llama.cpp naming conventions (full-precision conversions and BitNet ternary builds).