[stephrobert]

Summary

Eight audits ran on 2026-11-14 and none of them had a row in docs/audits.md. The register is a discipline, nothing fails when it is skipped, and skipping it is exactly how the same layers get audited twice while the same corners never do.

This adds the eight rows, and re-acknowledges one section of docs/limits.md that limits:check was refusing.

The rows

Each carries what was found and, in the last column, what was not looked at. That column is the one that chooses the next audit, and two of these rows already name a blind spot worth a future pass: no run against a real Incus on any of the three sibling ownership paths (#920, #932, #933), and internal/core/cloudinit never opened all day.

The pattern the day measured three times

Worth more than the eight rows, and the reason this is not just bookkeeping. In three lots, by three different authors, a contradicteur found the same defect: an attack test that refuses for the right reason by accident, because every row of its table shares a value that some other control refuses too.

• Attach does not ask whether the instance is ours, where Detach asks twice #909 drove one machine name that ownedNetwork also refused, so a guard reading the wrong name stayed invisible and only the accepting half caught it.
• falsify reads "no tests to run" as a guard that is not held #923 built a one-package module for its selftest, so "in this package" and "anywhere in the listing" were the same set, and the package filter could be deleted with the suite green.
• The cold twin of #909: a boot puts our machine on the operator's own network #932 drives networks that safeName and ownedNetwork both refuse, so the shape half of its guard is held by nothing at all.

The common cause is not carelessness: it is that the most obvious attack values fail every control at once. A guard that asks two questions needs a row that passes one and fails the other, in both directions. That sentence is in the register now; whether it belongs in CLAUDE.md as a rule is a separate decision and not taken here.

The limit

mise run limits:check was exiting 2 on main, and nothing caught it, because the task is deliberately not in prepush (its own help explains why: a red that interrupts daily work over documentation freshness teaches --no-verify, which disarms every other hook at once).

One section cited #851, closed 2026-11-13 and never re-acknowledged. The section was read: the limit holds. No client mints a request identity, the measurement over the 1088 recorded exchanges still stands, and the CLI refuses rather than pretending to be reproducible, which is what #851 asked for.

So the right treatment is a dated acknowledgement, not a rewrite. Only that one entry is re-dated, by hand. tools/docs/limits-acks.py refresh re-dates all 65 sections and would have claimed a pass nobody made.

Type of change

• A route added or changed
• Bug fix
• Refactor
• Documentation
• Chore / tooling
• Security / supply chain

Runs

run	verdict
mise run docs:check	green, on the rebase onto 806f5d3
mise run limits:check	green, exit 0, where main exits 2

Mode: nothing here touches code or a machine runtime, so no leg and no FEINT_VM run is owed.

No contradicteur read this, and that is deliberate

The project's rule is that a lot goes through contradicteurs before its commit. This one did not, and rather than leave the blank to be read as an oversight: it is a record of measurements taken in this session, plus one dated re-read of a limits section, and the two agent slots were held all day by the three code lots that needed them more. The rows cite the issues and the files they came from, so the claims are checkable without an agent.

Checklist

Always

• mise run check passes (no code changed; docs:check and limits:check are the gates that read these files)
• No new external Go dependency
• internal/core still knows no provider
• Nothing in the diff could be written identically for another provider
• mise run testplan was run; it names docs:check and limits:check for these paths, and both are green

When a model wrote a substantive part of this

• An Assisted-by: trailer names the tool and the model version
• No field name and no response shape: nothing a client can see changes

When a route is added or changed

N/A.

When behaviour a client can observe changes

N/A. docs/limits.md's prose is untouched; only its acknowledgement date moves.

When .github/workflows/ is touched

N/A.

When a machine runtime is involved

N/A.

Related issues

None. This closes no issue: docs/audits.md is a register and the limits acknowledgement is a dated read, neither of which an issue tracks.

🤖 Generated with Claude Code