chore(acp): bump dimcode, grok and kilo registry pins to probed versions

[kaizhou-lab]

Summary

Scheduled ACP Registry version sync. Three npx pins drifted since #960, each backed by a fresh serial ACP probe of the exact pinned version. This one is a single-file diff: no lock-derived test assertion embeds any of the three versions, and codebuddy — the only package whose version appears in a test — did not drift.

backend	package	old → new	initialize	session/new
dimcode	dimcode	0.3.26 → 0.3.27	ok (agentInfo version 0.3.27)	auth required (-32000, "Provider credentials are required")
grok	@xai-official/grok	1.0.16 → 1.0.17	ok (protocolVersion 1)	auth required (-32000, "no auth method id provided")
kilo	@kilocode/cli	7.5.6 → 7.5.9	ok (agentInfo Kilo 7.5.9)	succeeded unauthenticated

All three meet the release-lock criterion: initialize succeeds, and session/new either succeeds or returns a clearly classified authentication requirement. Probes ran serially with no inherited HOME or credentials. kilo opened a session advertising model, thought_level, and mode options; that catalog is evidence only and is not persisted (skill step 11 leaves those columns to the runtime). dimcode still self-reports agentInfo.title as "DimAgent" against a public listing of "DimCode" — a standing vendor quirk, not an AionCore defect.

Snapshot diff beyond the lock: a per-agent comparison of the raw CDN JSON against the previous snapshot shows every change is version churn — no distribution type was added or removed, and no entrypoint args or env changed for any of the 39 agents. Twelve non-lock agents also moved (claude-acp 0.73.0, cline 3.0.61, codex-acp 1.8.0, gemini-cli 0.58.0, factory-droid, plus binary artifacts for cursor, devin, harn, junie, kimi, opencode, and kilo's own binary channel). None of them is in our release lock, so all are report-only here; binary distributions in particular stay out of scope per the drift workflow's separate treatment of platform artifacts.

Derived-assertion scan: all three outgoing versions (0.3.26, 1.0.16, 7.5.6) were scanned across crates/**/*.rs with fixed-string matching before staging, and none has a single hit. registry_npx_lock.rs still pins codebuddy at 2.143.0, which is correct — codebuddy is unchanged in this snapshot. npx_cache_repair.rs keeps its own version literals: those are cache-path hash fixtures, not lock assertions.

The other 8 Registry-pinned packages (autohand, codebuddy, deepagents, dirac, glm-acp-agent, nova, pi, sigit) match the snapshot exactly. Package names and entrypoint args are unchanged for all 11. Drifted but not upgraded: none. mimo-code remains the one non-Registry builtin (no registry_json_id), excluded from drift reconciliation.

Registry snapshot

• Audit pinned to release tag v2026.09.02-fa32608 of agentclientprotocol/registry, fetched via the versioned CDN path for reproducibility.
• 39 ids in the raw snapshot: no newly listed and no delisted agents versus the baseline. (antigravity-acp remains listed and deferred as binary-only since 2026-08-21; fast-agent and minion-code remain listed but uvx-only and therefore out of scope.)

Validation

• just migration-check — pass
• just lint-fix (cargo fix + clippy --fix --workspace -D warnings) — clean
• just fmt — clean
• Local cargo nextest intentionally skipped, by standing policy for lock-only bumps (established 2026-08-11). The Test check on this PR is the authority for this change: the merge decision depends on CI rather than the local run, and this host's load only manufactures timeout-shaped test failures, which nothing in the local steps above is subject to.

Logging

No logging changes: this is a lock version bump only; existing startup/session error paths already identify a failing agent by backend.