Reference evaluator, stdlib round 1: 217 functions from the chapters — 412/610 evaluated, 409 agree, abstains 434 to 198; exact float text; findings F7-F10

[O6lvl4]

The stdlib long tail, first burn-down round (#10's standing ratchet).

Numbers over spec/wasm_cross + spec/programs (610 programs)

	before	after
evaluated by the reference	175	412
agree with the native target	171/172	409/412 — the 3 disagreements are all ADJUDICATED: F1 (named-arg order) and F10 ×2 (±0 min/max where 0.58.0 lags ratified ALS-T23), each under @ref-allow
evaluator faults	0	0
abstains (proofs/ref-abstain.toml, shrink-only)	434 / 154 classes	198 / 73 classes
stdlib functions implemented	17	217
λ_almd kernel	49/49 ×2	49/49 ×2

What lands

• ref/src/fmtfloat.rs — the evaluator's own float text: exact big-integer Dragon4 shortest digits (ALS-T13), to_fixed as round-half-even on the exact binary value (ALS-T9), and correctly rounded decimal→binary64 for ANY digit count/exponent (ALS-T2, denormals and saturation included). Held to the host formatter/parser by two oracle tests over ~6,000 samples — the host appears only as a test oracle.
• 217 stdlib functions written from ALS-S1..S6, T1..T24, C1..C10 and the fixtures those sections cite: T1's 25-code-point whitespace set, T5 full case mapping with the Final_Sigma rule, T6 termination aborts (division by zero, integer overflow, clamp/rotate/chunk messages), T8 parse grammars as pinned, T14/T16 mask and truncation doctrines, T23 zero-order min/max, T24 conversions, ADR-0006 first-err fallible callbacks in every HOF, in-place mutators.
• Semantics corrections the corpus forced: bindings are shared slots (a closure's var write is observed outside — closure_captured_list_mutation, sort_by_call_count; a shadowing let stays frozen — ref_gleam_capture_shadow); propagation carries its polarity so an Option-returning pure fn propagates none as none (C-211, pure_bang_propagation); fan.any/fan.settle; checker-style nominal retag of anonymous records with a unique field set.

New findings (PARSER-NOTES F7–F10) — normative prose vs executable evidence, before any implementation is judged

• F7 ALS-T16 says a negative count clamps 負→0; C-054's own fixture pins string.take(s, -1) = WHOLE (the as-usize doctrine). Prose and evidence disagree.
• F8 ALS-S3's one vacuous-truth sentence covers five different lifts: is_alphanumeric("") = false, is_digit is ASCII-only and empty-false, is_whitespace("") = TRUE, is_upper/lower are Python-style.
• F9 ALS-T8: int.parse trims the T1 Unicode whitespace set (T8 is silent); int.from_hex is not from_str_radix — lowercase 0x prefixes strip repeatedly (0x0x0x10 = 16), 0X does not, the sign may follow the prefix (0x-ff = -255).
• F10 release 0.58.0 still applies the retired first-argument ±0 rule in min/max — @ref-allow tracks both fixtures until the pin advances (the T23 chapter already declares its suite files red).

🤖 Generated with Claude Code