feat(server): separate default and maximum AdCP versions

[bokelley]

Summary

• separate the unversioned default served AdCP release from the maximum supported release
• preserve the immutable SDK-selected release across standard, custom-tool, discovery, resolver, deferred, and task handlers
• add an executable official-MCP compatibility-matrix lane proving unversioned 3.1 and explicit 3.2 callers can share one server

Validation

• npm run format:check
• npm run typecheck
• npm run build:lib
• node --test --test-timeout=180000 --test-force-exit test/lib/media-buy-lifecycle-compatibility.test.js (9/9)
• focused server and adapter suites (435 tests)

Local full-suite note

The broad local batched runner exposed two VM-only harness signals: an inherited DATABASE_URL changed a CLI scaffold expectation (the file passes 11/11 with it unset), and one unrelated batch reported cancellations with zero assertion failures. GitHub Actions runs in a clean, authoritative CI environment.

---

Open workspace in Conductor

[aao-secretariat]

Ladon verdict: Approve

Approve.

Checked: PR #2854 separates a server's default served AdCP release (defaultAdcpVersion) from its supported ceiling (adcpVersion) and threads an immutable servedAdcpVersion across handler, custom-tool, DecisioningPlatform, task-handoff, and resolver contexts plus a new responseEnhancer argument.

• Witness-preserving: unset default reproduces historical newest-advertised selection — no fabrication/re-shaping at seams.
• Construction validation is fail-closed with three gates; immutability and handoff threading implemented correctly.
• Minor changeset present and correctly classified for the src/lib/** surface change.

No Critical, High, or Medium findings. gated_paths is false and there is no no-auto-approve team match, so review_decision: REVIEW_REQUIRED does not gate. high_risk is false. No escalation triggers fire; falls through to row 9.

[aao-secretariat]

Ladon verdict: Approve

Approve — clean pass, no blocking or medium findings.

What I checked

• Reviewer surfaced no critical/high/medium/low findings this run.
• High-risk flag is true only via src/lib/protocols/mcp-modern.ts (modified) matching src/lib/protocols/**. Per the high-risk reading rules, a modified sensitive file with no medium-or-higher finding is presumed safe — not an escalation trigger on the flag alone.
• The delta refines MCP discovery: _meta.adcp_version version-signaling gated to major>=3 in SingleAgentClient/mcp-modern, plus a fail-open guard in create-adcp-server so externally-registered tools stay discoverable while the primary toolAvailableForRelease gate still applies. No wire-shape fabrication/normalization (witness-not-translator preserved).
• Changeset present (.changeset/quiet-versions-select.md), satisfying the changeset-vs-wire-impact requirement for src/lib/** edits.
• Tests exercise the delta.

Gate checks

• gated_paths: false → row 2 does not fire despite review_decision: REVIEW_REQUIRED.
• No no-auto-approve team match → row 7 does not fire.
• No (deleted) high-risk reasons → rows 3/6 do not fire.
• Prior decision was approve, so sticky escalation (row 6) is inapplicable.

No rows 1–8 fire → row 9 → approve.