fix(acp): reconcile hosted agent directory profiles

[LucasMoskun]

Summary

• reconcile buzz-acp identities into the replaceable kind 10100 agent directory after startup channel discovery
• publish actual channel names/IDs and effective response authorization while preserving unknown profile fields
• project the implicit runtime owner into the public allowlist, including the equivalent owner allowlist for owner-only mode
• retain an existing channel_add_policy, or require an explicit first-time BUZZ_ACP_CHANNEL_ADD_POLICY so startup never guesses or widens policy
• skip unchanged profiles and keep reconciliation failures non-fatal

Why

Desktop mention authorization now fails closed for remote agents missing current directory authorization. Legacy manually hosted agents can still be relay members and channel bots with valid kind 0 profiles, yet remain absent from autocomplete and receive no signed p tag. A live headless canary confirmed that a complete profile with the effective owner restores autocomplete, signed-recipient delivery, and harness wake without a Desktop restart.

Closes #5928.

Validation

• live canary: Desktop 0.5.14 autocomplete -> signed p tag -> relay -> buzz-acp wake
• cargo test -p buzz-acp (783 passed)
• cargo clippy -p buzz-acp --all-targets -- -D warnings
• just ci (full workspace/desktop/web/mobile gate passed; mobile: 1,399 tests)

[themiguelamador]

I found and fixed three issues in the new startup reconciliation path:

• Kind 10100 is replaceable, but the new event used the wall-clock second without advancing the observed head. A reconciliation in the same second as the current profile could therefore lose the NIP-16 event-ID tie-break and leave stale authorization/channel data authoritative.
• Buzz’s POST /events returns HTTP 200 for an ingest result even when its JSON says accepted:false. The code ignored that field and logged Published, so stale/rejected writes could be reported as successful. Publication now requires explicit acceptance.
• The operation is documented as non-fatal, but it performs up to three REST calls whose client retry budgets can collectively delay the subscribed main event loop for roughly two minutes during a partial HTTP outage. The whole best-effort reconciliation is now capped at ten seconds.

Fix: Complear/buzz branch review/pr-5929-fix, commit 8a837a5f28a69005c3a7c4db44b61d88b358acb4.

Verification:

• cargo test -p buzz-acp: 785 unit tests + 9 lifecycle integration tests passed
• focused profile reconciliation tests: 6 passed
• cargo clippy -p buzz-acp --all-targets -- -D warnings
• cargo fmt --check
• git diff --check

Requesting changes until the fix is incorporated.