Originally created by: 5dive-bot
TIER_CAL_BASELINE_US was 173000 — a measurement with no environment attached, taken on a runner
we no longer have. Every CI run for weeks printed RE-BASELINE THE CALIBRATION, the probe landing
29–33% below it, and the mechanism meant to absorb runner variance gave zero relief as a result:
the clamp floor is 100%, so a genuinely fast runner is crushed to the floor and its own probe saying
"I am fast" changes nothing.
Re-derived from CI to 119000, with the environment and the date recorded on the constant
per DIVE-2555 — a figure with no environment attached cannot be refuted by the next reading, only
silently disagreed with, which is exactly how 173000 survived.
This is a mis-tuning, not a stale number. Bumping the value without that reasoning would rot the
same way.
EFF_MS = BUDGET*1000*SCALE/100 with SCALE = clamp(raw, 100, 150). The floor at 100 means
EFF_BUDGET >= BUDGET for any baseline, so the set of harnesses that fit can only grow.
At the probe measured that day (~116476) the two constants are indistinguishable — both land on the
floor and both yield 1320s full / 300s core. What changes is only readings above 119000: a
genuinely slow runner, which 173000 crushed to the floor. Strictly a widening.
From this PR's own green unit-tests run:
241 harnesses, 225s wall-clock, budget 300s (75% of budget)
CALIBRATION (measured) 112674us/iter vs baseline 119000us/iter = 94% of a normal runner
applied 100% (clamp 100-150%) -> effective cap 300s
PROBE BRACKET (DIVE-2736) pre 112674us/iter -> post 113069us/iter (+0%)
No RE-BASELINE warning — 94% is inside tolerance where 65% was not. That is the false signal
going quiet, and it is sample [#1] of the concordant window the rest of DIVE-2736 (the SCALE_REF
split) is blocked on.
The row was parked on a wake condition of post-probe CI data — data that could not exist until the
branch landed. A pause whose wake condition depends on the thing the pause is blocking never
wakes. Compiled as a-park-that-waits-for-data-only-a-merge-can-produce-never-wakes.
full-sweep red, deliberatelyunit-tests — the budget gate, the one this change touches — is green. The full-* shards are
red, and they are not attributable to this change: main's own current head (4e87712) fails the
same jobs at the same step concurrently, without any of it, and the failing shards report 22–26%
of budget, so the calibration constant is irrelevant to them. Named separately as DIVE-2781
(selfcheck_mutation_e2e.sh post-#486, ui_views_e2e.sh pre-#486 — at least two distinct causes
being counted as one signal). No required check was bypassed.
Authored by dev3; committed under lodar <markounik@gmail.com> per this repo's house rule;
pushed on the root-side rail as 5dive-bot — dev3 holds no GitHub credential and could read CI
conclusions but never exit codes, so the log reads quoted above were done by agent-main.
Closes DIVE-2736 (the baseline half; the SCALE_REF split remains open on the row).
Ticket changed by: lodar