fix(color): preserve visible colors against hidden RGB bias

[phil-r]

Description:

Keep the snail navy instead of near-black when vectorizing its transparent PNG with 3 colors. Hidden black RGB shares a visible cluster, so all 3 palette slots survive and the slot recovery in #91 does not trigger.

Numeric requests now compensate shared palette centers for masked RGB contributions without reclustering, reordering layers or changing masks. Auto and binary behavior stay unchanged. This preserves the existing filtering and assignments rather than promising exact source colors.

This PR includes:

• The exact snail.png fixture and palette/render regressions for both profiles, plus alpha-boundary and independent per-cluster reference checks.
• Corrected Aftermath fill expectations, with its approved mask and geometry hashes unchanged and the original lower-level palette separately tested.

How to run:

docker compose --profile testing run --build --rm test \
  python -m pytest vectorizing/tests/test_color_numeric.py \
  vectorizing/tests/test_color_p3.py

Validation: 855 offline tests and all pre-commit hooks passed. test_integration.py (HTTP/S3/gallery integration) was excluded. Before/after captures confirm unchanged geometry and alpha; unaffected cases match exactly. No existing PNG baselines were replaced. Generated visual evidence remains local.

Local medians (M4 Pro, ARM64 Docker/Python 3.11, single-threaded, 5 fresh-process samples plus warmup):

Snail, 3 colors	Before	After
current	42 ms	46 ms
experimental	37 ms	40 ms

Mean-bias checking adds 22–33 ms on measured 1MP transparent inputs, including about 10–11% overhead on a populated palette whose output is unchanged. Peak RSS showed no meaningful increase. Timing excludes decoding, HTTP/S3 and rasterization; production concurrency was not tested.

Checklist (based on PR guidelines):

• I have self-reviewed this PR, according to these points
• This PR falls into maximum three of these categories
• I feel comfortable taking responsibility for merging this code to production
• The code is high quality, well-tested, follows styles guides (Frontend), clean, and well documented