fix(heartbeat): honor Ally concurrency 15

[allyblockcast]

Thinking Path

• Paperclip is the open source app people use to manage AI agents for work.
• Heartbeat dispatch converts queued work into bounded concurrent agent runs.
• Ally persists heartbeat.maxConcurrentRuns: 15, but the external-lifecycle operational ceiling is 8, so the runtime cannot honor the stored configuration.
• Raising admission before fixing queue-selection serialization could amplify overlapping scans, so this change is stacked on the strict dispatcher fix in fix(heartbeat): serialize queued-run dispatch instead of bypassing the lock (BLO-20396) #912.
• This pull request raises only the external operational ceiling to 16 and adds boundary, validation, and concurrent-claim regressions.
• The benefit is that Ally can use its configured 15 slots without letting concurrent starter calls claim a 16th or allowing unsupported API values to persist.

Linked Issues or Issue Description

Paperclip issue: https://paperclip.blockcast.net/BLO/issues/BLO-20742

Related prerequisite: #912. Prior eight-slot ceiling change: #885.

Bug report fields:

• What happened: Ally stores maxConcurrentRuns: 15, while external-lifecycle admission clamps it to 8.
• Expected behavior: a valid persisted value of 15 produces 15 available slots, with values above the supported validator maximum rejected before persistence and external admission bounded by a hard operational ceiling.
• Steps to reproduce: configure an opencode_k8s agent with concurrencyEnabled: true and maxConcurrentRuns: 15; under backlog, observe no more than 8 active runs before this change.
• Deployment mode: authenticated Paperclip Kubernetes deployment.

What Changed

• Raise EXTERNAL_LIFECYCLE_SLOT_CAPACITY from 8 to 16 so Ally's configured 15 is effective.
• Prove resolveExternalLifecycleConcurrency returns 15 for the configured value and clamps 16/50 to 16.
• Prove create/update agent validation accepts 15 and rejects HEARTBEAT_POLICY_MAX_CONCURRENT_MAX + 1 before persistence.
• Add an embedded-Postgres regression that invokes 20 concurrent dispatch starters and asserts exactly 15 runs are claimed while five remain queued.

Verification

• pnpm exec vitest run packages/shared/src/adapter-types.test.ts server/src/__tests__/heartbeat-external-lifecycle-concurrency-flag.test.ts server/src/__tests__/heartbeat-start-lock.test.ts --reporter=dot (32/32 passed)
• pnpm exec vitest run server/src/__tests__/heartbeat-dispatch-priority-sort.test.ts -t 'claims at most 15 available slots across concurrent dispatch attempts' --reporter=dot (1/1 passed)
• pnpm --filter @paperclipai/shared typecheck (passed)
• pnpm --filter @paperclipai/server typecheck (passed)
• Pre-rollout capacity at 2026-08-02 02:43 UTC: nine Paperclip nodes expose 368 CPU and ~661 GiB allocatable, with ~221 CPU and ~361 GiB requested. Six observed Ally pods used 7.376 CPU / 3.161 GiB; each requests 1.1 CPU / 2.25 GiB, so scaling from six to 15 adds 9.9 CPU / 20.25 GiB of requests. No Ally FailedScheduling event was observed.

Risks

• This raises the maximum blast radius of a misbehaving external-lifecycle agent. The per-agent config remains a separate gate, so other agents do not change unless configured above 8.
• Do not merge or deploy ahead of fix(heartbeat): serialize queued-run dispatch instead of bypassing the lock (BLO-20396) #912. This PR is intentionally stacked on that branch so strict queue-selection serialization lands first.
• A recent Ally pod had a transient prompt-secret FailedMount; it was not a resource-capacity failure, but post-rollout verification must still watch FailedScheduling, OOMKilled, eviction, start-lock errors, queue age, and review failure rate.
• Rollback is the one-line operational ceiling revert; setting concurrencyEnabled: false immediately restores one-run containment.

For core feature work, check ROADMAP.md first and discuss it in #dev before opening the PR. Feature PRs that overlap with planned core work may need to be redirected. See CONTRIBUTING.md.

Model Used

• OpenAI GPT-5.6-sol (openai/gpt-5.6-sol) via OpenCode, with reasoning, repository editing, test execution, GitHub, Paperclip, Prometheus, and Kubernetes read-only tools. Context window size was not exposed by the runtime.

Checklist

• I have included a thinking path that traces from project context to this change
• I have specified the model used (with version and capability details)
• I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work
• I have searched GitHub for duplicate or related PRs and linked them above
• I described the issue in-PR following the bug template
• I have run tests locally and they pass
• I have added or updated tests where applicable
• This change has no UI impact
• No documentation change is required beyond the PR rollout contract
• I have considered and documented risks above
• All Paperclip CI gates are green
• Greptile is 5/5 with no open P2s, recommendations, or follow-ups
• I will address all Greptile and reviewer comments before requesting merge

[allyblockcast]

🔗 Paperclip issue: BLO-20742

[allyblockcast]

🔗 Paperclip issue: BLO-20742

[allyblockcast]

@ally please review head 39ed31f with focus on the 15-slot concurrent admission invariant, validator/runtime consistency, and compatibility with stacked prerequisite #912.

[allyblockcast]

Superseded by an exact-head restack on current #912; this branch used a stale managed-local pointer and GitHub correctly reported it conflicted.