fix(cu): consolidate target identity validation

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.

[hqhq1025]

Summary

Consolidates and supersedes #930, #931, #932, and #933.

• binds coordinate mutations to the actionable source element identity;
• validates Electron page/document identity for every pointer mutation before fallback;
• fails closed for unknown process classification;
• consumes the opaque element_token protocol from feat(cua-driver): add stable AX node identity tokens trycua/cua#2210;
• requires fresh tokens for semantic click/set_value;
• routes actionable coordinate clicks through fresh AX token dispatch, never pixel fallback;
• keeps native type fail-closed until a focused-node token API exists;
• preserves idempotent write/readback behavior.

Verification

• Core full suite
• Runtime: 1943 tests, 7 skipped, 0 fail
• Computer Use: 144/144
• CI will rerun on the updated head

Real-machine validation

Using a local build of trycua/cua#2210 rebased onto cua-driver 0.8.1, OpenAI gpt-5.6-sol completed:

• AppKit AX set_value
• AppKit AX click_element
• multi-step set_value → fresh-token click_element, including recovery from a physical-input intervention
• process restart recovery: stale target returned target_missing, then the model rediscovered the new PID/window and succeeded

All successful mutations used AX dispatch and recorded zero pixel dispatch.

Upstream dependency

Keep draft until trycua/cua#2210 is reviewed, merged, and released. Its fork workflows currently require maintainer approval to run.

[hqhq1025]

Adversarial review removed the remaining attribute-based node substitution. Latest head 464d58b1 dispatches the exact opaque token from the consumed observation, rejects caller identity mismatches, never re-snapshots and guesses a replacement node before mutation, and preserves driver same-node changed/readback_value as internal evidence without rewriting a fresh observation. Candidate driver f15f8d87 passed real gpt-5.6-sol AppKit set_value, click, multi-step set_value→click, and restart recovery with zero pixel dispatch. This remains Draft until upstream #2210 is merged and released, then the released artifact/version/hash/provenance must be pinned and the same qualification rerun.

[hqhq1025]

2026-07-20 dependency audit: keep this PR Draft; it is not mergeable yet. Upstream trycua/cua#2210 remains OPEN and DIRTY after maintainer review found two P1 blockers (empty Linux/Windows snapshots can panic; destructive token consumers can discard the exact retained-node binding) and one P2 blocker (cross-platform snapshot_id schema drift). Newer cua-driver prereleases up to v0.9.0 exist, but #2210 has not merged or shipped in a release, and Maka is still pinned to cua-driver-rs-v0.7.1-maka.2. The old green CI only validates the downstream adapter against its candidate protocol; it does not satisfy the release/provenance gate. A dedicated worker is now fixing #2210. Once upstream merges and releases, this branch must be rebased onto current main, the released artifact/version/hash must be pinned, and real AX qualification rerun before converting to Ready.

[hqhq1025]

Upstream blocker update: trycua/cua#2210 has now been rebased to current upstream/main and all three maintainer findings are fixed on head c40481bd. The upstream PR is MERGEABLE (no longer DIRTY); current GitHub state is BLOCKED only by REVIEW_REQUIRED. Validation passed: core 293, macOS 152, Linux 16, Windows 32, protocol/schema 6, full cargo fmt, diff check, and conflict-free merge-tree. This downstream PR remains Draft until #2210 is merged and included in an official cua-driver release. After release, Maka still needs to pin the released version/archive/binary hashes, rebase this branch onto current main, and rerun real AX + restart qualification before converting to Ready.