fix(planner): FP16-baseline consistency for quality/tier (#4) + cut 0.6.2

[Sahil170595]

Summary

Completes the fix for #4. 0.6.1 stopped quality_tier returning unknown for llama3.1-8b, but only patched quality_tier() — estimate() and predict() stayed on the old FP16-baseline chain. Result: the FP16 config reported quality 0.5 with provenance unknown (ranking the highest-precision option below Q2_K at 0.59), while its tier said negligible. Quality and tier disagreed.

This unifies all three methods behind one _fp16_baseline resolver, then cuts 0.6.2.

The fix

QualityModel.predict(), estimate(), and quality_tier() now share _fp16_baseline(model, family) — one resolution chain:

1. measured FP16 baseline (fp16_baselines)
2. explicit model|FP16 lookup entry
3. the model's own highest-precision measured quant (TR125 uses Q8_0 as the llama3.1-8b baseline — 8B FP16 = 16 GB doesn't fit the RTX 4080, so no FP16 row exists)
4. family mean (off-registry models)

Quality, provenance, and tier can no longer diverge. No FP16 number is fabricated in fitted_models.json — the anchor is the honest one TR125 already uses.

Before → after (plan --model-size 8b --hardware "RTX 4090 24GB" --request-rate 2.0 --json)

quant	before	after
FP16	quality=0.5, prov=unknown, tier=negligible (inconsistent)	quality=0.635, prov=estimated, tier=negligible
Q8_0/Q6_K/…	unchanged (measured)	unchanged (measured)

Tests

• New regression test_estimate_and_tier_share_baseline: estimate() and quality_tier() resolve the same baseline; FP16 is estimated (not the 0.5/unknown prior) and not ranked below Q2_K.
• Full suite: 488 passed, ruff clean.

Release

• 0.6.1 → 0.6.2 (pyproject.toml, __init__.py)
• CHANGELOG.md [0.6.2] entry
• CLAUDE.md version + test-count refresh (483 → 488, correcting pre-existing drift)