Locust is not CICADA, and four things this session got wrong on the way there

[syncytium2]

Locust is not CICADA, and four things this session got wrong on the way there

The session record, written before compaction. Straight into docs/handoffs/
rather than the root: nothing here is half-done, and the root HANDOFF.md
belongs to the 2026-08-27 session and is still honest — its five PRs (#304,
#292, #270, #53, #50) were each verified OPEN at writing time. Replacing it
would destroy a live signal.

The finding it exists for: locust's derivation chain is validated only at its
last link. gen_ref_cicada.m builds the 1e-9 parity fixture by running
interface2's own generate_sce_cicada, so nothing anywhere compares anything to
the Cossart source; both repos assert "faithful port" and neither tests it. It
is also a documented partial port — the per-cell transient-detection step is
skipped — and interface2 PARKED that function in f55643bf for over-detecting on
this preparation's long SLOW transients.

Established so nobody re-derives it: the port came from their CODE, not the
paper (sce_stats_utils.py exists at the cited path with both named functions,
and the rasterdur parameter confirms the duration model it replaces); the
citation is correct (Zenodo 10041434, v1.0.3, five authors, 20 July 2020,
CC-BY-4.0), checked deliberately because the lit shelf holds DeepCINAC by the
same five authors in the same year; and the duration diagnosis has been in
interface2's methods doc since July with the mechanism and the number.

It also carries four self-corrections and the pattern they share — a claim true
of one context asserted about another — because the pattern is more useful to
the next session than the fixes. The worst was a todo proposing a bench run at
onset_field="peak", which would have moved the detector in the direction the
project deliberately came from; retracted the same day.

And the hazard worth the most: a worktree's pytest imports the PRIMARY
checkout's src, because the venv holds one editable install. It fails toward
green, and it had already put "1454 passed, 16 skipped" into two PR messages
when the true figure was 1498 passed, 3 skipped, 1 xfailed.

Review scope is stated in the file: a single-pass claim check against git, the
files and the live page, not an eleven-role murderboard. Said plainly because a
handoff claiming more review than it had is the defect this repo files
incidents about.

🤖 Generated with Claude Code