feat(transform): adaptive chars-per-token for the profitability gate (opt-in)

[dhruvraajeev]

The gate weighs image tokens (exact, from the 28px patch grid) against text tokens (chars / CHARS_PER_TOKEN). That CPT constant is hand-tuned: 4 for reminders/tool_results, 2.0 for slab/history. docs/ADAPTIVE_CPT_PLAN.md measured real density near ~1.5 on dense content, so the gate over-estimates how cheap text is and declines compressions that would have won.

This learns CPT from telemetry pxpipe already writes. bucket_chars, baseline_tokens and image_pixels have been logged for a while and nothing read them; a 6-bucket least-squares fit over those rows solves a marginal token rate per bucket (CPT = 1/slope). No new dependencies.

src/core/cpt-fit.ts the fit. Pure math, no fs, safe in the Workers build.
src/cpt-store.ts Node-only. Reads events.jsonl, writes cpt-state.jsonl,
hands the gate a resolver.

Fails closed everywhere, which is the part worth reviewing. Min 20 samples, min 8 bucket appearances, plausible band 0.8-6.0, condition-number check, non-positive slope, throwing resolver — each rejects that one bucket back to its baked constant and records why in cpt-state.jsonl. A wrong learned rate costs money; a missing one is just today's behavior. Precedence is explicit host charsPerToken -> learned -> constant. Workers keeps the constants outright.

Image cost is subtracted on the patch grid from anthropic-vision.ts, not the retired pixels/750 figure. Pages are 1568x728 and both edges are multiples of 28, so pixels/784 reproduces the patch count exactly (56x26 = 1456).

Off unless PXPIPE_ADAPTIVE_CPT=1; refits are throttled and never block a request. Verified: the fit recovers CPTs it was generated from within 5% including under noise, and a learned rate flips a real gate verdict both ways. Savings on production traffic are NOT measured — that needs a shadow-mode A/B, which is why this is opt-in. Reproduce: npx tsx eval/adaptive-cpt/demo-cpt.mts