feat(coexistcheck): warn on a default agent the suppressed identity no longer serves

[arcaven]

After adopt without --rewrite-agent, a repo's default agent can name an identity that adopt has just suppressed, so a session offers no such agent and coexist-check still reports a clean preflight. This adds the missing layer-5 warning, so an operator who skipped the dry run learns about it at the next check instead of in a session. It also stops the preflight from calling a run with warnings "clean".

Item 2 of #95; item 3 merged in #194, so this completes #95. Additive except for one changed output line: a run that raised warnings now ends on preflight passed with N warning(s) where it used to end on preflight clean. Nothing parses that line in this repo; the docs line is updated.

The architect ruled WARN, with the banner fix (valid until 2026-10-22).

What changed:

• checkDefaultAgent (internal/coexistcheck/coexistcheck.go) reads the effective default agent, settings.local.json over settings.json. It warns only when the agent carries the foreign <pack>: prefix and an identity of that pack is suppressed in the repo. The warning names the agent and both remedies: sideshow adopt <pack> --rewrite-agent and sideshow activate <pack>.
• PassVerdict(warns) returns preflight clean only for zero warnings and zero errors, else preflight passed with N warning(s). coexist-check uses it for its last line.
• The adopt dry run's last line (adopt.go:215) can follow printed WARNs: the dry run prints the preflight findings, then the four plan steps, then this line, so a WARN is still on screen above it. It uses the same rule: preflight passed with N warning(s) and every plan step resolves: the real run would proceed. dryRunPreflight now returns the warning count it printed.
• docs/repo-bindings-enablement.md describes both endings.

Pins (red in 1c96b21, green in the fix commit):

• Foreign default in settings.json, identity suppressed: WARN naming both remedies, and the end-to-end coexist-check run ends on passed with 1 warning(s).
• Foreign agent in settings.local.json over a clean settings.json: fires.
• Clean local agent over a foreign settings.json: does not fire.
• Foreign default whose identity is enabled: no finding. Bound (non-foreign) agent beside a suppressed identity: no finding.
• A run with no warning still prints preflight clean, for coexist-check and for the adopt dry run.
• Adopt dry run with a WARN: counted verdict.

Mutants, each compiled and killed by a named test: suppression ignored (..BesideEnabledIdentityDoesNotWarn); committed file read over local (..InLocalSettingsWarns, ..CleanLocalAgentOverForeignCommittedDoesNotWarn); any agent fires (..BoundAgentBesideSuppressedIdentityDoesNotWarn); verdict always clean (the three banner tests); verdict never clean (the two no-warning controls); adopt drops the count (TestAdopt_DryRunCountsWarningsInItsVerdict); coexist-check ignores warnings (its banner tests); finding raised as INFO (the three WARN tests).

Gates, each run on its own: build, vet, go test -count=1 ./..., golangci-lint (0 issues), gofumpt -l (empty).

Closes #95

[arcavenai]

Approving. After adopt without --rewrite-agent, a repo can be left with a default agent that no longer resolves. coexist-check now warns about that instead of calling the run clean, and a run that raised warnings now says so on its last line. On main, the same 2-WARN repo ended on preflight clean: enable/adopt may proceed; at head it ends on preflight passed with 2 warning(s): enable/adopt may proceed.

What I ran. I built both binaries (main 33fa181, head b4123c4) and used a scratch HOME and CLAUDE_CONFIG_DIR with the network denied. The fixtures are git init repos with no commits. Every go test ran with -count=1.

The new check. I ran 20 settings layouts through Run with a scratch test, since deleted.

case	head
foreign agent in settings.json, identity suppressed by settings.local.json	1 WARN naming the agent, the identity and both remedies
foreign agent in the local file over a clean committed one	WARN
clean local agent over a foreign committed one	none
local "agent": "", or no local key, over a foreign committed one	falls back to the committed file; WARN
suppression in the committed file against a user-scope enable, agent there	WARN
own form vsdd-orch or vsdd:orch, unprefixed orch, another pack's prefix, the lookalike vsdd-factory-x:orch	none
foreign prefix, identity enabled or never enabled	none
no default agent	none
malformed settings.local.json or settings.json	Run returns parse <file>: unexpected end of JSON input from the census, before this check, the same as on main; no panic and no WARN
a non-string "agent" in the local file	the local file is ignored and the committed value is used

End to end, coexist-check gives these results:

• Dangling agent: 1 WARN, ending preflight passed with 1 warning(s).
• A fresh wave-state.yaml alone: 2 WARNs, ending ... 2 warning(s).
• Both together: 3 WARNs, ending ... 3 warning(s).
• A clean repo: still ends preflight clean.

PassVerdict is preflight clean only at zero warnings, and it is reached only when Refuse() is false, so a run with an ERROR never prints it. It prints N warning(s) for any N above 0, which is the form the ruling names.

Red, mutants, and consumers.

• Red at 1c96b21: the 5 positive tests fail and the 5 controls pass. All 10 pass at head, and ./... passes.
• Mutants: each of the 8 listed is killed by the tests the body names. So are three I added: an off-by-one count, an empty local agent overriding the committed one, and the adopt count increment dropped.
• Consumers: git grep 'preflight clean' finds 3 lines on main, the positive control. At head it finds only PassVerdict, the docs line and the new tests' assertions. No script, workflow, justfile or other doc parses the line.
• CI: b4123c4 passes.

Non-blocking:

• False WARN after a re-enable. The WARN text says "no such agent resolves", and that is false when an identity of the pack is listed as suppressed but the pack is still effectively enabled. Two cases do this:

  • a user-scope enable, a committed false, and a local true, so the identity is re-enabled locally;
  • a second marketplace's identity of the same plugin enabled beside the suppressed one.

  In both, the run already refuses on same-repo-dual-enable, so the verdict line is unaffected. Only the extra WARN in the refused output is wrong. Returning early when view.EffectivelyEnabled holds an identity of the pack would remove it.

• Uncovered mutant. Counting INFO findings as warnings in dryRunPreflight (case foreign.Warn, foreign.Info:) survives, because neither dry-run fixture raises an INFO. A control with one INFO finding would pin the count to WARN only.

• "Directly under". The body says the dry run's last line sits directly under the printed WARNs. In the test's output the four-step plan block sits between them, so the line is a few lines below, not adjacent. The reason for counting holds either way: both are on the same screen of one run.

Seat: reviewer / reviewer