Golden-value regression tests for the record layer (P2.5)

[remislp]

Every package in the stack reads its records through this code, so a change
here reaches SCALCS, EKDIST and HJCFIT at once. Pinning it at the bottom rather
than in HJCFIT is the same work with three times the coverage.

What is pinned

Five cases over the two shipped records, each fixing resolved-interval counts,
period counts, burst counts and openings, mean burst length, mean Popen within
bursts, and the histogram binning.

record	tres	tcrit	resolved	bursts	mean length
experimental	25 µs	4 ms	3232	63	98.629089 ms
experimental	100 µs	4 ms	2408	62	100.030402 ms
experimental	500 µs	10 ms	1812	56	111.190894 ms
simulated	500 µs	4 ms	33377	9270	111.686132 ms
simulated	1 ms	10 ms	23095	7021	149.705263 ms

The resolutions were chosen so the dead time actually bites. The simulated
record's shortest interval is 125 µs, so a golden case at 25 or 100 µs merges
nothing — it would pin the concatenation logic while exercising none of it.
One of the invariants asserts this, so a future case cannot quietly test
nothing.

Alongside the exact numbers, invariants that hold for any record at any
resolution: periods alternate strictly, every burst has an odd interval count,
burst openings account for every open period, no within-burst gap exceeds
tcrit, the histogram counts every interval, the sign of tcrit is ignored.
Those keep meaning something if the example files are ever replaced.

Two invariants were wrong, and the failures were worth following

I wrote both as obvious truths; both failed on the experimental record at
larger tres. Neither was a defect — but relaxing the tolerance would have
buried two real facts about the layer.

Total time is not conserved outright. impose_resolution starts at the
first interval that is both resolvable and usable, because anything earlier has
a start that cannot be located, and the time before it is dropped — 75.8 µs of
a 102.7 s record at 100 µs. Exactly the observed difference. The invariant now
accounts for that head explicitly.

Resolved intervals are not all at least tres long. Two documented
exceptions carry FLAG_UNUSABLE: the negative sentinel for an unfinished final
opening, and the short shut emitted when a record ends on a brief shutting that
terminates an open run. The invariant is now about usable intervals, which is
what a consumer relies on.

83 tests here; 345 passing overall.

🤖 Generated with Claude Code