fix: four of the #280 follow-ups: mint credentials, listing dedupe, ensure identity, launch_image_arn

[laithalsaadoon]

What and why

Four of #280's "Follow-ups found during the burn-down", each in its own commit with its failing-first test. Based on main at cae2ec7, which carries #536 (the per-crate semver gates, merged as def2815), so those gates judge this change too.

• A credential failure during a token mint is ERR_CREDENTIALS (c3046d4). ProxyAuth::token_for reclassified every minter failure as WireKind::AuthTokenMint, so a mint the control plane refused for the caller's identity was ERR_RETRYABLE (exit 3) with a message promising "the identical request may succeed". A minter failure of kind Credentials keeps it, with the minter's reason as its source; a throttle and every other mint failure stay retryable.
• ControlPlane::list_microvms lists each VM once (7148700). The service can return one VM on two pages while VMs change state (docs/PLATFORM.md, measured 2026-10-01 by fix(conformance): CLI-8 names the run's VM from its history, and counts a listing's VMs once #500), and the listing concatenated pages, so ls --remote and both bindings' ControlPlane.list could list a VM twice. A repeat keeps the VM's first position and the entry read last, the newer reading of its state.
• An ensured image's name covers its create-only fields (499ae33). The identity hash left out identity repair, inherit_workdir, the hook timeouts and the log group and stream, so a reuse could return an image built without a field the caller asked for (from feat(cli): build --reuse, run's build and agent-up go through core's ensure_image, and the aws s3 cp upload is gone (#258) #454). prepare names the image for create_identity_hash of the exact create request it sends; the stream's tag moves to microvms-ensure-image/2, so every ensured image, the agent recipes' included, is renamed once. The public pinned_identity_hash and image_identity_hash keep their signatures and answer the identity with the create request's defaults. Tags stay out of the name.
• Both bindings expose launch_image_arn (707f785). Python's Sandbox.launch_image_arn property and Node's Sandbox.launchImageArn() read core's Sandbox::launch_image_arn (fix: Sandbox::run resolves a bare image name to its ARN, for every surface (#253) #401), with a launch-image-arn row in verify/parity/capabilities.toml naming all four surfaces (the CLI's is run --image).

Changelog: changelog.d/280.fixed.md (three entries) and changelog.d/280.added.md.

Size: 424 changed lines of product code against origin/main, past the soft cap, because the owner asked for these four #280 boxes as one branch. The change can be split: each box is its own commit, so review can take it commit by commit, and the ensure rename (499ae33) is most of the lines.

Verified offline only. Nothing here was exercised against AWS; the dedupe follows the 2026-10-01 measurement, and the mint path's credential class is asserted against a fake minter.

Evidence

• Main moved to 8d87a3b (nightly chore(deps): nightly rollup 2026-10-09 #538, docs docs: correct cli.md's microvm.toml section and envelope key lists against the code #539) after the rebase. b69c0ee merges into it without conflicts, and TMPDIR=<scratch>/t mise run -c check on that merged tree exits 0 in 310 s (cold build).

• cargo fmt --all -- --check

• cargo clippy --all-targets -- -D warnings (through mise run check)

• cargo test --all (through mise run check; the workspace's Rust tests also pass with cargo test --workspace --exclude microvms-py --exclude microvms-js --no-fail-fast)

• TMPDIR=<scratch>/t mise run -c check at the final commit (b69c0ee, rebased onto cae2ec7): exit 0 in 110 s (warm; the cold run before the guards re-proof, at 10c21e7, exited 0 in 301 s). mise run semver:check exits 0 too: all five comparisons against 0.11.0 pass (196 checks: 196 pass, 58 skip each). Before the rebase, a run failed lint on a clippy::type_complexity in the new ensure test, fixed by a type alias in 499ae33.

• Both binding suites, built the way bindings:py and bindings:js build them: pytest 461 passed, 1 skipped; node --test 349 passed, 0 failed.

• ./tools/generate-py-stubs.py and mise run dts regenerated microvms.pyi and index.d.ts; stubs:check and dts:check pass in check.

Failing first, each watched red before its fix:

• session::proxy::tests::a_credential_failure_during_a_mint_is_err_credentials_not_retryable: left: Retryable, right: Credentials.
• control::microvm::tests::a_vm_the_fleet_listing_meets_on_two_pages_is_listed_once: left: ["mvm-a", "mvm-moved", "mvm-moved", "mvm-b"].
• control::ensure::tests::each_create_only_field_names_a_different_image: repair_guest_identity is create-only, so it is identity, both names task-48daea63e51c.
• test_launch_image_arn_names_the_image_the_launch_sent: AttributeError: 'microvms.Sandbox' object has no attribute 'launch_image_arn'; node's launchImageArn names the image the launch sent: TypeError: sandbox.launchImageArn is not a function.

Parity

• No public name changed on any surface (core, CLI commands, microvms.pyi, index.d.ts). Python gains Sandbox.launch_image_arn and Node Sandbox.launchImageArn(); core's public paths are unchanged (core-api:check passes).
• verify/parity/capabilities.toml updated: the launch-image-arn row names core, run --image, Sandbox.launch_image_arn and Sandbox.launchImageArn. mise run parity:check passes.

The three image-name cases in verify/parity/cases/ take the renamed images (parity-task-178447f17502, agent-vm-claude-code-cb158b8e2c7d, agent-vm-claude-code-codex-937344ac9c1b); core, the CLI and both bindings' runners answer them.

Guards

mise run fail-to-pass -- --jobs 4 --emit parity-followups against cae2ec73845c, the merge base with origin/main, wrote six entries into verify/guards/faults/parity-followups.toml with their patches:

fired: f2p-a-created-image-carries-every-field-of-the-request (9.2 s)
fired: f2p-each-create-only-field-names-a-different-image (9.1 s)
fired: f2p-a-vm-the-fleet-listing-meets-on-two-pages-is-listed-once (9.4 s)
fired: f2p-a-credential-failure-during-a-mint-is-err-credentials (2.9 s)
fired: f2p-test-launch-image-arn-names-the-image-the-launch-sent (15.7 s)
fired: f2p-launchimagearn-names-the-image-the-launch-sent (25.3 s)

control::ensure::tests::an_ensure_from_a_create_request_keeps_its_fields changed too (its direct ensure takes the same log group and identity repair as the create it compares to) and passes with the fix taken out, so it proves nothing about the fix and has no entry; the two ensure entries above guard the rename.

Follow-ups

• The cost report's cycles default differs: cost --cycles defaults to 1 and the bindings' suspend_resume_cycles to 0. Left open: it is a product decision for the owner (from test: every surface default is held to core's through the parity table (#300) #457), and stays on tracking: core parity across the CLI, Python and TypeScript #280's checklist.
• core+bindings: TCP tunnel and HTTP port-forward in the SDKs, with identity= on launch #263 and core: move the CLI's probe-then-register attach and the budget gate into core #270 stay open: their boxes wait for live runs that drive the bindings' tunnel handles and the import probe.
• No live check runs a binding's import_record / importRecord against a real VM (from feat: registering a record another machine wrote is core's (#270) #470); stays on tracking: core parity across the CLI, Python and TypeScript #280's checklist.
• The renamed ensured images leave the old /1-named images in an account; nothing deletes them. A caller who wants them gone deletes them by name.

Platform claims

• Not applicable: this changes no claim about AWS behavior.

Scope

• This is not an orchestrator, a fork implementation, or AgentCore parity work
  (see docs/STRATEGY.md).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.