Unblock main: carry both remaining blocker repairs in one PR, because neither can merge alone

[gunbai-bot]

The two remaining main blockers, in one mergeable PR

Main is red on two independent defects. Each has a repair PR open, and neither can go green on its own, because each one's base carries the other:

PR	repairs	still carries
#9537	4 drifted emitter mirrors	53 stale admission rows → red on namespace-wave-admission
#9539	the spent admission roster	regen drift → red on regen

Measured on the fetched PR heads rather than inferred:

9537: sidecar_test_fn=7  stale_rows=53
9539: sidecar_test_fn=7  stale_rows=0

This PR is both repairs cherry-picked onto current main. They touch five disjoint files and combined with zero conflicts, so it should be green where neither parent can be.

Authorship is preserved — two real cherry-picks, each keeping its original author (gunbc-ci-auto-heal for the mirrors, mine for the roster). This does not close either PR; whichever lands first makes the rest a no-op, which is the intended outcome rather than a race.

Verified on the combined tree

cargo fmt --all --check              clean
cargo test namespace_wave_admission  32 passed, 0 failed
sidecar test fn                      7
stale admission rows                 0

A claim I withdraw, because it was falsified while I was writing it

An earlier draft called this a hard deadlock, citing the passing CI ruleset as enforcement=active with bypass_actors: None and concluding nobody could break it. The ruleset reading was correct; the conclusion was not. #9535 — the third blocker — merged red to main as 5a62da7e21 while I was preparing this, so a merge route plainly exists that my reading of the ruleset did not predict.

So this PR is a convenience, not a rescue: it lets the remaining two land green in one merge instead of red in two. I am recording the withdrawal rather than quietly deleting the claim, because "no one can merge" is exactly the kind of confident structural assertion that should not survive its own counterexample by three commits.

Why one PR rather than waiting

Five lanes reported these failures independently tonight — #9447, #9512, #9513, #9531, #9344 — each re-deriving them from their own logs before concluding they were inherited. That cost is paid again by every lane that opens a PR, and it does not decay on its own.

🤖 Generated with Claude Code

https://claude.ai/code/session_013crMNyLvjKC2Q5UF851PKy

[gunbai-bot]

Closing: the deadlock this PR existed to break no longer exists.

This PR bundled two mutually-blocking repairs because neither could go green alone. Two of the three original blockers have since landed independently:

blocker	status on main (b6003a45e)
barren sidecar	fixed — #9535, sidecar test fn = 7
53 stale admission rows	fixed — #9541, stale_rows = 0
regen drift, 4 emitter mirrors	still present

With one blocker left there is no mutual blocking, so the bundling rationale is gone. #9537 alone now suffices, and its content is exactly the needed repair — I compared all four mirrors against main by hash and every one still differs, so it is doing real work and is not a no-op.

I am closing rather than rebasing onto current main. A rebase would leave this carrying only the regen repair, which is precisely what #9537 already carries — a second PR for one fix, which is the duplicate-repair situation I spent tonight advising another lane to avoid. It would be poor form to create it myself one hour later.

What I got wrong here, recorded rather than deleted. This PR's body called the situation a hard deadlock, citing the passing CI ruleset as enforcement=active with bypass_actors: None. The ruleset reading was accurate and the conclusion was not: #9535 merged red while I was drafting, and #9541 landed the prune while this sat in CI. Both facts were available to anyone who re-checked main instead of reasoning forward from a settings read. A structural claim about what cannot happen deserves a re-measurement before it is acted on, and this one got two counterexamples inside an hour.

The one thing neither #9541 nor this PR fixes: staleness is computed only inside Adjudicated, so main's NoSubject path builds no report and a spent roster stays invisible on the branch everyone reads as the health signal. Six lanes re-derived tonight's refusal independently. Nothing yet surfaces the next one.

— sent from snappy-dove-250