test(comms): direct coverage for resolve.go's handle-resolution contracts (RIG-3536)

[rigel-mintaka]

This PR is part of a stack containing 3 PRs:

1. main
2. test(e2e): marker-routed multi-turn scripts + observer-scoped fixture clients (RIG-3528) #1029
3. "test(comms): direct coverage for resolve.go's handle-resolution contracts (RIG-3536)" (this PR)
4. test(e2e): multi-tenant comms visibility transport proof (RIG-3531) #1042

T8 of the RIG-3473 record. resolve.go had no dedicated test file: its contracts
were pinned only where other files happened to cross them.

The load-bearing case is the vantage-probe closure. resolveVisibleAgentHandle
maps a real-but-invisible agent to the SAME NOT_FOUND an unknown handle gets, and
that INDISTINGUISHABILITY is the security property -- a distinguishable miss is
an existence oracle. So the test asserts it as an equivalence: each submitted
spelling is substituted out and the two messages must be byte-identical, the
resolved id must not appear, and both must map to CodeNotFound. A paired positive
stops a resolver that returns NOT_FOUND for everything from satisfying it, and a
fixture guard fails loudly if the invisible agent ever becomes visible, so the
assertion cannot go vacuous.

The deliberate ASYMMETRY is pinned in the same test: resolveAgentHandle is
owner-namespaced but NOT viewer-scoped, so it still resolves that agent. Both
halves in one place, because the pair is the contract.

Also: submitted-order preservation (ids that do not sort with their handles, so a
map-iteration bug yields a wrong permutation rather than an accidental pass); the
same bare handle under two owners resolving differently per caller; an unknown
owner qualifier naming the SUBMITTED handle; notFoundHandle passing a
non-not-found error through by IDENTITY; and empty input short-circuiting before
any store call (driven against a nil store, so it can only pass if the guard
fires).

Batch atomicity is NOT re-proven here. resolveHandles delegates it to
store.AccountsByHandles, already covered by TestAccountsByHandlesAtomicMissNamesAll
-- so this tier asserts only the pass-through: the sentinel survives, no partial
result rides along, and the message equals the store's own verbatim.

Test-only; resolve.go is unmodified. Mutation-verified: leaking the resolved id
into the visible-miss, ignoring the visibility result, and always re-keying in
notFoundHandle each fail with a message naming the defect. gofmt/vet/vet -tags
pgtest clean; -race pgtest run ok against a live Postgres, with the new tests
confirmed RUN rather than skipped.

Ledger-impact: none

Refs RIG-3536

Co-authored-by: Matt Wilkinson matt@rigel.build

[linear-code]

RIG-3536

[github-actions]

Compass engineering docs preview: https://compass-comms-rig3536-resolv.compass-eng-docs.pages.dev

Deployed from compass-comms/rig3536-resolve-coverage at 003ceb5.

[trunk-io]

Stacked PR 1042 was cancelled: a user cancelled it.