[Polichinel]

reports/post_mortems/2026-10-20_release_v1150.md, following the convention of
2026-08-03_releases_v190_to_v1110.md.

Why this one is worth reading

The v1.11.0 post-mortem named the spine running through everything it reported:

"a written record that was true when written and stopped being true without anything failing."

It also opened by noting that nine releases had shipped without a post-mortem, and that the
convention had recorded itself failing before failing nine more times.

Both halves recurred. Three of the four releases since are unwritten and now never will be. And
the spine produced five fresh instances in eight days, all in C-343 — the fifth found inside
the register entry created to track the first, including in its Trigger: field, the line an
operator is meant to act on.

The conclusion the document records is not "try harder". It is that nothing we run detects this.
Every instance in this window was found by a human asking an unprompted question, or by an author
re-reading their own work.

Scope

Release	v1.15.0 — published and deployed 2026-10-20
Commits	8 in v1.14.0..main
PRs	#500, #502, #503, #504, #506, #507
Issues opened	#501 (open), #505 (auto-opened, auto-closed), #508 (open)

What it covers

• The measurement that started it. The 2026-10-16 run took 5h50m; the zarr export alone was 5h07m — 88% of the pipeline, against ADR-041's recorded ~57 minutes. A 5× regression, four months old, that nothing reported. Found because a human pasted a log.
• A release whose wheel contains no new code. git diff --name-only v1.14.0 origin/development -- src/ is empty. The release exists to carry a script and corrected guides to a tag-deployed server, and the document says so plainly rather than letting the version number imply otherwise.
• The spine, round by round, including how each round fixed the previous round's blind spot — and docs: a count is a copy too — four places still said deploying is three steps (C-343) #507's finding that four prior rounds audited content and never cardinality.
• An incident section for a mistake made in this window: reporting that the instrumentation would reach production "automatically" when the server runs a tag that did not contain it. C-343 displaced one layer up, committed three days after the same author wrote round four.

It records what went well, not only what failed

• Block 4 of the deploy procedure executed successfully for the first time since it was written — on 2026-10-16 its first command died on the ~-expansion defect, so docs: ~ after sudo expands as the caller — six verification commands read the wrong home (C-343) #498's fix is now confirmed by execution rather than by reading.
• The The zarr export is 88% of the pipeline and ~5x its last recorded cost — measure before optimising #501 falsification criterion was committed to the issue before any production figure existed, with the synthetic baseline as the stated contrast.
• Three guards were sabotaged and confirmed red rather than assumed to work.
• The clean-venv publish check reported No matching distribution found — and the broken thing was the check (Python 3.10 venv against requires-python >=3.11), not the release. A less careful reading would have raised a false alarm on a good publish.

The gaps it names and does not fix

1. Nothing compares durations across runs — check_health.py has 26 freshness references and zero duration references. Flagged, not built: one datapoint does not justify a trend watcher.
2. Nothing compares what the server runs against what the repo expects, at any layer. Story 3: Run the deploy gates somewhere other than one laptop (C-341) #424's gate would not have caught this window's variant.
3. Audit dimensions are chosen ad hoc, which is why five rounds were needed rather than one.
4. docs/sources/shdi.md says "we do not redistribute GDL data" — the served store broadcasts four GDL values verbatim #508 — docs/sources/shdi.md says "we do not redistribute GDL data" while the served store broadcasts four GDL values verbatim. The same spine in a different domain.

Verification

Check	Result
bash docs/validate_docs.sh	PASSED
All 14 register-reading test files	41 passed

Documentation-only; no code, no tests, no behaviour. Every figure is cited to a file rather than
recalled, and the release count was corrected against git tag before committing — the first
draft said "thirteen consecutive releases" from reasoning rather than counting, which would have
been the document's own spine inside the document.

Refs #501, #508.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EfeuULbPTqrx4qtNki5dy9