alternate port: anthropic/claude-sonnet-5 — 37/37 gated effects, $17.2118, retrieval instead of context

[staccDOTsol]

alternate port: anthropic/claude-sonnet-5 — $17.2118 all-in (DHH did not run this model)

One of five frontier models given DHH's exact challenge (same plan.md, same
Python reference) with one change: the 396k-token reference was retrieved
per-ask through a holographic-memory layer instead of re-sent in context.

• 37/37 effects, each gated on compiling AND rendering styled ANSI animation
• $17.2118 total, provider-billed, every ask on the attached LEDGER.jsonl
• 4.3x measured vs carrying the corpus (real cold/warm sends, not assumed cache rates)
• watch the effects run from this binary: https://ttfx.awesomemcp.fun/show/sonnet

On the pure-math subset, we ran YOUR goldens against sonnet's code on glibc
(the platform class that generated them): 15,014/15,015 samples bit-exact
(99.993%) — all 12 implemented named easings at 1001/1001, and all three
cubic-bezier goldens attempted via a transcribed make_easing Newton solver
(the sibling ports skipped beziers): two of three bit-exact, the third at
1000/1001 with the single miss 1 ulp off (1.4e-17) — an LLVM lowering artifact
we reproduced but couldn't close, reported rather than trimmed. The 354-case
parity suite remains this repo's bar and is not claimed here.

This adds alternates/anthropic_claude-sonnet-5/: the same challenge, run on the cost
axis, with a receipt for every ask.

honesty section (condensed from the run's live during-mortem)

The harness was improved WHILE this ran, and every dollar of that is in the
numbers above, not edited out: a render gate added mid-run dropped and re-bought
already-paid effects; three reasoning-budget truncation bugs bought ~$7 of
unusable responses before diagnosis; early asks shipped 3x-fat retrieval chunks;
restarts re-bought cores. Projections use the blended all-waste rate (marginal
rates run ~5x lower). The "without retrieval" comparison is MEASURED, not
modeled — real cold+warm full-corpus sends at provider-billed prices, which
found NO default cache discount for Anthropic or DeepSeek-via-OpenRouter. DHH's
side counts only his published successes; ours counts every failure. The
finished Rust port in this repo was quarantined from the models throughout —
they saw only plan.md and the Python source. Full ledger of every ask attached
(LEDGER.jsonl: tokens, provider-billed USD, timestamps).

Close it if exhibits don't belong here — the branch and the live page stand on
their own either way.