6b48edb2c added assertion (1c), the span slack, with a per-configuration constant of 0.25 ms per leaf record (3 ms under either sanitizer) capped at half the leaf. Its own evidence is four green runs per lane.
It reds intermittently. A fresh review of #1536, working on a branch whose merge base already contains 6b48edb2c, observed CHECK( span_slack <= span_bound ) failing on decode.video at 0.00256913 <= 0.00075 — 3.4x the bound — and saw it again once under an unrelated mutation. Four further runs of that case alone were green. The box was at load average 70-125.
It is pre-existing and independent of #1536: that row does not touch (1c), keeps its constants exactly as they landed, and the observation is on a tree whose only difference from the merge base is the residue work.
Worth pairing with what that change's own text records: the span slack is an ABSOLUTE bound on a quantity whose cost is instrumentation, which is why it needed a 12x per-configuration constant in the first place, and its note says in terms that "a NORMALISED bound would be better, and there is no normaliser". Record::instrument_seconds now exists and is one.
FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]
6b48edb2cadded assertion (1c), the span slack, with a per-configuration constant of 0.25 ms per leaf record (3 ms under either sanitizer) capped at half the leaf. Its own evidence is four green runs per lane.It reds intermittently. A fresh review of #1536, working on a branch whose merge base already contains
6b48edb2c, observedCHECK( span_slack <= span_bound )failing ondecode.videoat0.00256913 <= 0.00075— 3.4x the bound — and saw it again once under an unrelated mutation. Four further runs of that case alone were green. The box was at load average 70-125.It is pre-existing and independent of #1536: that row does not touch (1c), keeps its constants exactly as they landed, and the observation is on a tree whose only difference from the merge base is the residue work.
Worth pairing with what that change's own text records: the span slack is an ABSOLUTE bound on a quantity whose cost is instrumentation, which is why it needed a 12x per-configuration constant in the first place, and its note says in terms that "a NORMALISED bound would be better, and there is no normaliser".
Record::instrument_secondsnow exists and is one.FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]