docs(register): IR-90 — the disk red that wore a product costume, and the two terms IR-85 was missing

A full disk on hfenduleam did not fail a build. It failed a test, in our own
code, with our own refusal string: `sync refused: bundle failed` at
spt-daemon/tests/sync.rs:198, with "Out of diskspace" sitting in the fourth
nested clause behind a git exit code. C: was at 0.018 GiB free of 1862.02 at
that instant. Every layer behaved correctly, and the report reads as a sync
regression at the sha under test — which is worse than the two disk faces
already banked, because LNK1318 and rustc I/O route a reader to "the box is
sick" while this one routes to a code owner. It landed mid-discriminator with
an old sha and a new sha side by side, where the cheapest reading is a head
regression that does not exist.

The remedy is one guard in two placements, and it REFUSES rather than runs: a
free-space floor at rig start, and free space appended to a shelled-out
git/store error before it becomes a product refusal string. Both, because a
floor read once at rig start would have passed the #209 lane at 136.32 GiB and
that run still ended at 12.27 — 221 test binaries took 88 GiB in ten minutes.
A start-only guard is a guard against yesterday's disk.

Two things that looked like fallout are recorded as NOT fallout, so the entry
cannot overclaim: serve list's SERVE_UNCONFIRMED is the 0.67.0 resident broker
image lacking servehost.rs (doyle, from source), and the 4-of-7 quic DEGRADED
is 10.2 days old. Even the pump stall is claimed as concurrent, not caused —
one sample of a "last tick N s ago" counter cannot separate still-ticking from
ticked-once.

IR-85 gains the two terms it did not name when it landed at d32d5c4c. ci.yml's
classify step at :49 emits code=true for every non-pull_request event, so a
push to main runs unit on BOTH runners and the docs-only skip is
pull_request-only — doyle's finding, named as his own error, and the reason run
34337797758 completed FAILURE at the 40-minute wall on a docs-only delta over a
golden-green sha. A near-full volume joins its environment term, deliberately
unquantified because the window that would have apportioned it was void.

What the voided discriminator did establish is negative and stated as such: the
slowdown reproduces at 04e32c8c, 3.2x to 4.7x on all five tests, so head growth
is not NECESSARY. It measures nothing about how much the environment explains,
and arm 1 stays open.

Register-only. No workflow, no code, no REQ tag, no ledger row.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WAw9XAcTR8oSwzYPsqZoY2

Co-authored by: hertz