test(connscale): emit five band-less diagnostic fields so a moved ratio can be attributed (BACKLOG #1366)

[wshallwshall]

Built by Builder 2. Landed by the Lander. One commit, four files, +390/-9. Zero files under messagefoundry/ or .github/.

What it does

The connscale smoke emits five band-less diagnostic fields to the step summary on every run. The existing ratio says that something moved; these say which. Drain-tail and reload-probe separate only on drain_seconds against reload_seconds; contention separates from probe-cost only on the FD probe's tick counts. All five already existed on the record and died with the test job, which uploads no artifacts.

This is directly relevant to BACKLOG #1357, whose connscale monotonicity assertion has been failing intermittently across unrelated PRs all evening with no way to attribute the movement.

The design call worth a reviewer's eye

This is a second renderer, not a parameter on the existing one.

Only two connscale metrics carry an SLO band; these five have none. Emitting prior, band-floor and margin for a band-less field would print a threshold computed from whichever reading happened to precede it -- manufactured precision that reads as measured in a job summary. So the new table carries no band, no threshold and no verdict, and says so in its own preamble.

Two defects the author's own verification caught

A duplicate test name silently deleted #236's cap test. A new test was given the name of an existing one. Python replaced the earlier definition with no error and no warning: the file held 29 def test_ and ran 28. It was caught only by counting definitions against collected. There is now a guard that fails the file if any two tests share a name.

Two of the author's own tests were hollow, and mutation proved it. One iterated the constant it was testing, so renaming a field out of the shipped set survived. The other asserted a substring that also appears in the heading, so deleting the sentence it meant to pin survived. Both are now pinned against literals.

Six mutants, all killed, each by a distinct red set.

Lander verification

commits vs origin/main                 1
behind origin/main                     0
git merge-tree --write-tree vs main    exit 0 (clean)
git diff --shortstat origin/main...    4 files changed, 390 insertions(+), 9 deletions(-)
files under messagefoundry/ or .github/    0
'+## 1366' in the BACKLOG diff             1

Exact match to the handover. I also re-ran the author's own duplicate-name control rather than taking it:

tests/test_connscale_smoke.py               defs 9    unique 9    OK
tests/test_connscale_empty_claims_per_msg.py  defs 32   unique 32   OK

The +## 1366 check is there deliberately: PR 623 was refused by the required a PR that implements BACKLOG #N must update BACKLOG.md context for exactly the missing row, and the author folded that lesson into this handover rather than needing to be told twice.

Not armed

Required approvals are 0, so arming means landing unread. This is diagnostics rather than a fix for a live defect, so nothing on main is costing anything while it waits for a reader.

Generated with Claude Code