fix: explain unavailable deployment acquisition

[alongubkin]

Summary

• return bounded reasons when explicitly requested deployments cannot be acquired
• classify SQLite acquisition misses in the same transaction as lease selection and mutation
• fail fast on permanent acquisition mismatches while retaining bounded retries for contention and deferral
• preserve compatibility with older managers and downstream deployment stores

Validation

• cargo check -p alien-manager --features sqlite,openapi
• cargo test -p alien-manager --features sqlite explicit_acquire_ --test sqlite_store_tests -- --nocapture
• cargo test -p alien-deployment explicit_acquire_mismatch_fails_without_retrying --lib -- --nocapture
• pnpm run generate:manager-rust-sdk
• generated TypeScript SDK build and compatibility tests

Linear: ALIEN-948

Refreshed against current main. Re-ran two SQLite acquisition classification tests and three generated client compatibility tests. Regenerated with pnpm run generate:manager-rust-sdk and pnpm run generate:manager-api; fresh schemas also reconcile existing external-binding schema drift. Live isolated API/manager proof distinguishes statusMismatch from a held lease and confirms successful lease release. The CLI transport HTTP regression executes all 60 contention attempts over 118 seconds and reports the held-lease explanation; a live CLI contention run now also passes. The test acquired and renewed an owned AWS deployment setup lease before launching the real CLI in supported standalone-manager mode against an isolated manager and API over local HTTPS. After the bounded wait, the CLI explained that another operation still held the lease and that this command had not acquired it. Holder release returned HTTP 200 / success=true. Earlier hosted-mode runs over the public tunnel ended with HTTP 503 and remain inconclusive for that routing path. Timeout text also distinguishes deferral and older managers without reasons. Normal platform SDK generation was run to align newly exposed reason fields.

Platform SDK

client-sdks/platform is reset to main in this PR. This repo has no generator for the platform spec, so the earlier platform SDK additions (UnacquiredDeployment, DeploymentAcquireUnavailableReason, the new SyncAcquireResponse fields) are dropped here. They come back on the next platform spec sync. Nothing in this repo depends on them: the CLI transport reads acquisition reasons only through the manager SDK (alien_manager_api::types).

Manager artifacts after merging main: alien-manager-schema-exporter output is byte-identical to crates/alien-manager/openapi.json; the generate:manager-rust-sdk down-convert/fix steps and pnpm run generate:manager-api (Speakeasy 1.680.11) produce no changes besides generation metadata.

[greptile-apps]

Confidence Score: 5/5

[Critical risk] Updates the public API contract for deployment acquisition.

The PR appears safe to merge; no outstanding correctness, compatibility, security, or repository-rule failures were identified.

Summary

This PR adds bounded explanations for explicitly requested deployments that cannot be acquired and surfaces those explanations through deployment clients.

• Classifies acquisition misses atomically with SQLite lease selection and mutation.
• Fails immediately for permanent request mismatches while retaining bounded retries for contention, deferral, and older managers.
• Adds optional acquisition-reason models to the Manager API and generated clients without breaking the existing sync request model.
• The previously reported AgentSyncRequest compatibility issue is resolved.

Diagram

sequenceDiagram
    participant C as Deployment client
    participant M as Manager sync route
    participant S as Deployment store
    C->>M: Acquire explicit deployment IDs
    M->>S: Select, lease, and classify in one transaction
    alt Deployment acquired
        S-->>M: Acquired deployment
        M-->>C: Deployment and execution claim
    else Temporary unavailability
        S-->>M: Contended or deferred
        M-->>C: Optional bounded reason
        C->>C: Retry within bounded wait
    else Permanent mismatch
        S-->>M: Model, platform, status, or mode mismatch
        M-->>C: Optional bounded reason
        C->>C: Fail without further retries
    end

Loading

_{Reviews (3) · Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." · Reviewed by Greptile}

[alongubkin]

@greptileai review

[alongubkin]

@greptileai review