benchd: seal the timed-token tolerance figures only coarsened

[GumbiiDigital]

Summary

The timed-token tolerance path added in the 2026-09-28 snapshot seals engine-chosen values verbatim:

• token_mismatch_count, token_mismatch_first_step, token_mismatch_near_tie_count (and the second-choice count and max relative gap), in the flat metrics and in every paired_legs row, including refused runs;
• the TIMED-DIVERGENCE-NOT-A-NEAR-TIE text, which interpolates the first diverging position and the reference gap at full precision, and the TIMED-DIVERGENCE-OVER-TOLERANCE text, which interpolates the candidate's mismatch count.

The sealed record goes back to the participant, whose engine has seen the hidden prompt and chooses which tokens it commits. Exact values therefore form a small read-back channel (several bits per pair).

Change

• Sealed record: the counts seal as a power-of-two bucket (0 stays 0), the max second-choice gap at 2 decimals, and token_mismatch_first_step not at all (absent on the flat metrics, null on a pair row, so the row's key set is unchanged). The judge still decides on the exact in-memory values.
• Refusal texts: NOT-A-NEAR-TIE names no position and rounds the gap to 2 decimals. OVER-TOLERANCE states the allowance benchd computes from the fixture instead of the candidate's count. The tolerance refusal seals first_failing_step as null.
• Logging: exact values go to stderr for operators.

2 files: crates/benchd/src/official.rs, crates/benchd/src/score.rs.

Tests

• New the_sealed_record_holds_no_exact_token_counts: count 13, first step 45 and gap 0.123456 seal as 16, no step, and 0.12, on the flat metrics and a pair row.
• Updated the key-pin bytes, the additive key set and the two refusal-text assertions.
• cargo fmt --check, cargo clippy --workspace --all-targets -D warnings and all benchd tests pass (655, 1 ignored).
• A few bench-runner process-spawn tests (not touched here) fail intermittently under parallel runs and pass with --test-threads=1.

🤖 Generated with Claude Code