borg's enrolment baseline is three populations, and only one is a finding
- Type: bug (Track T — host enrolment / triage)
- Found: 2026-09-11, enrolling borg as the watcher host after seven's retirement.
What happened
T stopped publishing: seven was retired into plexus at 16:29Z on 2026-09-11
and plexus has not run since 2026-08-30 (it is the dev box; the owner's
2026-09-01 directive says T deliberately does not run there). borg was
re-enrolled the same evening and published its first full as
tstate(borg): d4f170a4e7fb RED (full) BASELINE:16 red
A host with no prior job map has no baseline to diff against, so all 16 are recorded as NEW-RED. That is correct behaviour and a trap for the reader: the report is indistinguishable from a tree that broke 16 ways this evening.
The split, measured
| n | population | evidence | status |
|---|---|---|---|
| 7 | wasm32 RUNNER-ABSENT |
report's own toolchain: … wasmtime=ABSENT; every failure text reads wasmtime not found, so target 'wasm32' was NOT RUN |
fixed — wasmtime 48.0.1 installed to ~/.local/bin, matching the version seven reported. Gone from the next run's red list. |
| 3 | no 32-bit dev libs | gcc -m32 cannot link here: cannot find Scrt1.o, cannot find -lgcc, bits/libc-header-start.h: No such file. Hits test-emit-obj#00/#01/#02 |
needs root — apt install gcc-multilib libc6-dev-i386, with the owner |
| 2 | same cause, other direction | test-c-abi-mixed-link / test-record-abi-mixed-link SKIP i386 on the recipes' own gcc -m32 probe — scored passlike, invisible in a GREEN |
same fix |
| ? | residual | lib-test#src:tools/crtl_reachability.py, test-core#c_crtl_wait.c, test-core#c_pty_family.c, demos#00, tools-devtest#00, test-selfcompile-odiff#00 |
the only finding, and not yet separated from the row below |
The row that makes the residual hard: the toolchain INVERTED
seven's last report and borg's first, side by side:
| seven (2026-09-11T16:28Z) | borg (2026-09-11T19:51Z) | |
|---|---|---|
| gcc | 15.2.0 | 13.3.0 |
| qemu | 10.2.1 | 8.2.2 |
| git | 2.53.0 | 2.43.0 |
| kernel | 7.0.0-31 | 7.0.0-29 |
The host running T is now two toolchain generations OLDER than the box every
current expectation was tuned against — and meta/hosts.json shows seven itself
was on gcc 13.3.0 until 2026-09-05, so the fleet has already moved once in the
other direction. Any differential probe against the gcc oracle now answers
about a different gcc than it did last week, and nothing in the harness says so.
crtl_reachability's failure text is exactly the shape that predicts:
conflicting types for typedef 'int64_t' — a repeated typedef must name the same type, from headers resolved out of the host /usr/include. That is
track-t.md's own host-coupling rule, whose oracle is ldd, not the shape of
the assertion.
So the residual cannot be triaged as tree defects until the two host gaps are closed and one more full has run. Ordering matters: fixing the host first costs one sweep; triaging first costs a wrong diagnosis per job.
Why this is filed as a bug and not a note
Two of the three populations were produced by conditions
tools/twatch-setup.sh called READY. That half is fixed in the same change as
this ticket — the script now derives its runner list from
twatch.RUNNER_BINARIES (it named four qemu binaries; the constant names seven,
and the three it omitted included wasmtime) and probes gcc -m32 with the
recipes' own command. What is left here is the triage, which is not automatable:
deciding which of six residual reds are the tree.
Next step
apt install gcc-multilib libc6-dev-i386(root; with the owner).- One full tier on borg, then diff its red set against this baseline.
- Whatever is still red is the finding — triage it as host-coupling first
(
lddthe built test before reading the assertion), not as a regression.