[codetalcott]

Release m0serve 1.15.0 and m0 0.12.0. It carries the fixes from the 2026-10-10 review (#615) and the two issues reported against 1.14.0 (#611, #612).

What is in the branch

The bump (388f666f):

• Versions: pyproject.toml, uv.lock, M0SERVE_VERSION in cli.mojo, the m0 wheel's __version__, and QUICKSTART's echo.
• The changelog: the [Unreleased] entries move under [1.15.0] — 2026-10-11 with an introduction, and the Fixed list is grouped by component.

This is a minor release because two changes touch the served contract. The introduction and the Changed entries name both, as 1.13.0 named its port change:

• A request header whose name holds an underscore is dropped before a WSGI application sees it, or an ASGI application under --blocking-threads (the CVE-2015-0219 class).
• --replay-frames or M0_REPLAY_FRAMES above 4096 is refused at startup with exit 78.

Two re-pointed sabotage anchors. In the run, new: writes over what is there and image: --tag is ignored read NOT APPLICABLE: #615's P8 and P9 had moved their lines, and #615 re-ran the m0-wheel harness but not this one, whose image rules need docker. Both are re-pointed in the last commit, re-run alone, and caught.

Bench artifacts (e3985bc4): bench-linux-conclusions' three Linux artifacts and the pool probe. render-bench-docs reports them already current.

Not needed for this release:

• Bench re-records: the six macOS kinds were recorded at 1.14.0 and are next due at 1.16.0.
• A re-soak: both soaks read 1.14.0, and they go stale at 1.17.0.
• The pool A/Bs: neither the pool's records nor its park path changed since 1.14.0. What did change is the drain (loop/shutdown.mojo, SIGTERM drain waits out its 5 s deadline on a keep-alive connection holding the start of a next request #611) and the shim's publish numbering.

The pre-release run (every gate in docs/RELEASING.md)

Stage	Gates	Result
A	fuzz-request-long, soak-apps, smoke-wheel, four browser gates, sabotage-differential, test-postgres-server	all green
C	autobahn; stress-pool (20/20 rounds and 40/40 holds in all three wake modes); probe-mojo-image; sabotage-mojo-image	all green
Q	probe-pool-fairness: in job order with the barrier on, and the convoy and starvation visible with it off. probe-pool: pooled p99 of 0.2 / 0.3 / 0.2 ms at slow 0 / 1 / 2. stress-asgi: 30/30 in both modes. bench-linux-conclusions: every conclusion holds, spreads 0.7–1.2 %.	all green
S	sabotage-notes-login 15/15, sabotage-blobs 31/31, sabotage-outbox-cap 4/4, sabotage-m0-wheel 26/26, sabotage-host 68/68 plus parallel 1/1	all caught
D	sabotage-scaffold: 72 rules outside image:, then each of the 12 image: rules alone, with a builder prune and an fstrim between	84/84 caught

The quiet stage ran before S and D because the machine went quiet early in the run.

After the merge

Tags m0-v0.12.0, then v1.15.0, at the merge commit. Each release run pauses at its PyPI environment for the owner's approval. Approve pypi-m0 first, then pypi: the site deploys when the Release run completes, so this order keeps it from ever describing an m0 the index lacks.

🤖 Generated with Claude Code