fix(android): serialize direct GATT connection admission

[ODevStudio]

Summary

Part 2/5 of decentespresso/decaid#871; related #26.

• Serialize Android direct connection establishment with a per-plugin FIFO. Established peers keep independent reads, writes and notifications.
• Cancel queued requests without creating a GATT. Hold a failed/cancelled native owner until its exact GATT closes.
• Fence stale callbacks and adapter-off work by owner identity and epoch.
• Preserve existing deadlines, cooldowns and autoConnect=true behavior.

This hardens recovery; it does not establish the cause of the original field timeout. No global GATT mutex, adapter reset, heartbeat, or scale-protocol change.

Simplification

Removed the separate scenario runner/script and duplicate cases: 245 fewer lines. Unique ownership tests now run directly in the existing Android test job; plugin-level recovery and healthy-peer regressions remain.

Verification

At 6184659: analysis clean; Flutter 136 passed / 13 skipped; Android 63 passed. Android counts now include individually reported ownership cases previously grouped under one runner.
Final cross-platform CI passed.

Deferred: affected Android + DE1 + original-scale acceptance in decentespresso/decaid#877. PR remains draft.

Integration

PR #28 supplies native cleanup integration and end-to-end attempt cancellation. A caller timeout is not teardown confirmation. Unconfirmed native cleanup must remain blocked rather than force-close a healthy peer.

Design and scope | Acceptance evidence

Contributor Note

AI-assisted implementation and review. Maintainer review and affected-device acceptance remain open; nothing is merged.

[ODevStudio]

The interrupted Decaid #871 work has now been split into bounded, linked work packages. This draft was recovered and is the reusable candidate; no duplicate fork PR has been opened.

Tracking overview: decentespresso/decaid#871

• [871 · 1/5] Establish the current BLE baseline, correlate failures, and select the recovery fix decentespresso/decaid#875 — current baseline, correlated diagnostics and solution decision.
• [decaid#871 · 2/5] Harden the direct-connect admission state machine and deterministic regressions #26 — admission state-machine contract, cancellation/deadline ownership and deterministic regressions.
• [decaid#871 · 3/5] Integrate native admission with GATT teardown, cancellation and Android build tests #27 — native GATT lifecycle integration, Android compilation and actual plugin integration tests.
• [871 · 4/5] Integrate the selected BLE recovery fix with cancellation, retry ownership and an exact fork pin decentespresso/decaid#876 — Decaid timeout/retry/scan integration and exact fork pin.
• [871 · 5/5] Final BLE recovery regression, affected-Android validation and documentation sign-off decentespresso/decaid#877 — final software/affected-device validation, documentation and rollback.

Important baseline update: decentespresso/decaid#867 now preserves a healthy original-scale connection for displayOff after the #874 findings (inspected head 0b3da3942c55940d40939fc66cdee6d0d5f847f8). Baseline and candidate tests must include the same corrected scale-power behavior; explicit disconnect mode is separate.

At inspected head ac8427a466f122a072756639a4a850ef4cafef51, the reported JVM scenarios and current green checks are useful evidence, but native Android build/integration coverage and affected-device results still need explicit verification under #27 and the final acceptance issue. Cross-device overlap remains a candidate cause, not a hardware-proven explanation for the initial timeout. The decision package may prefer a narrower or different fix.

This comment only links the plan; it does not change the draft status, approve the implementation, or claim new test execution. Dependency work remains in this fork, not upstream; no dart_js change is currently justified.