fix(compression): let explicit interrupts cancel safely

[suparious]

What does this PR do?

Makes interrupt-protected context compression cancellable by an explicit user or lifecycle stop, without weakening protection against ordinary incoming messages, voice interjections, or active-turn redirects.

The reproduced CLI failure was:

1. context compression held the session compression lease while generating its handoff summary;
2. the user pressed Ctrl+C;
3. interrupt protection kept the summary worker alive after the frontend's grace period;
4. the CLI accepted another turn while that worker still held the lease;
5. the next persistence operation failed closed with session storage could not be written.

This change:

• separates explicit hard cancellation from ordinary interrupt/redirect state with a dedicated threading.Event;
• introduces AuxiliaryExplicitCancellation, an attempt-local frozen-cause signal that cannot be confused with provider/plugin/OS InterruptedError;
• isolates only the synchronous provider callback and stream aggregation in a bounded daemon worker during protected compression, allowing the owning transaction to unwind promptly without closing or evicting process-shared OpenAI-compatible, Codex, Anthropic, or Bedrock clients;
• propagates caller ContextVar state plus provider progress and interrupt-protection hooks into the isolated call;
• atomically linearizes Codex timeout cleanup against explicit cancellation, preventing a mixed outcome where the owner reports cancellation after the shared-client timeout path has already won;
• propagates hard cancellation through child agents and explicit stop surfaces: CLI Ctrl+C/signals, gateway/API/ACP stop and shutdown, SSE disconnect/cancellation, inactivity timeouts, cron, lifecycle cancellation, and foreground/background delegation cancellation;
• keeps old/synthetic/third-party agent compatibility through statically feature-detected hard_interrupt() with a legacy interrupt(message=None) fallback;
• serializes hard-cancel admission against compression commit admission with CompressionCommitFence, so cancellation either wins before mutation or waits for an already-admitted commit to finish;
• aborts before session rotation or late DB commit, restores in-place transcript mutations and an explicit snapshot of compressor summary/provenance/count/savings/cooldown/error/fallback/progress/telemetry state, and releases the heartbeat and compression lease;
• snapshots and verifies the exact raw durable cooldown row under the session lease (including expired/null states), and surfaces compensating-write failures instead of reporting a false no-op;
• preserves the established failure/fallback behavior for unrelated InterruptedError instances;
• adds deterministic cross-thread regressions for provider-call isolation, shared-client safety, frozen cancellation cause, hard-versus-soft interruption, compatibility fallback, pre-commit races, full rollback, lock release, recovered append, and successful retry.

A cancelled provider daemon can outlive its owner until the request's existing provider timeout. It cannot access transcript/commit state or retain the session lease, and no shared client is closed or evicted to stop it.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

How Has This Been Tested?

Post-rebase verification against current NousResearch/hermes-agent:main:

• canonical affected suite: 1,178 passed, 0 failed across 63 isolated test files after rebasing onto the upstream test-pruning sweep;
• compression concurrency/rollback file: 28 passed, 0 failed;
• TUI protocol/compatibility file: 39 passed, 0 failed;
• Ruff: passed for all 25 changed files;
• Windows footgun scan: passed for all 25 changed files;
• git diff --check: passed;
• added-line secret/injection/debug scan: passed.

A repository-wide run also exercised over 50,000 tests. It exposed eight patch-related compatibility failures; all were corrected and are covered by the green affected suites above. Remaining repository-wide failures were unrelated profile opt-out/cache test-environment failures in unchanged skill/Honcho paths.

Test Configuration

• OS: Linux
• Python: 3.12.13 via uv
• Test runner: scripts/run_tests.sh (per-file isolated subprocesses)

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing tests covering the changed behavior pass locally
• Any dependent changes have been merged and published in downstream modules

Documentation and downstream dependency changes are not applicable: this is internal interrupt, provider-call isolation, and compression-transaction behavior.

[alt-glitch]

This was generated by AI during triage.

Related to merged #49928: that patch masks ordinary incoming-message interrupts during compression; this patch adds a distinct fence-linearized explicit-stop cancellation path across frontends.

[teknium1]

Thanks for separating explicit cancellation from ordinary incoming-message interruption. The underlying gap remains on current main: CLI Ctrl+C uses interrupt() at cli.py:15457-15466, compression enters aux_interrupt_protection() at agent/context_compressor.py:3591-3599, and the Codex stream suppresses interruption while that protection is active at agent/auxiliary_client.py:1183-1188.

No concrete blocking implementation defect was confirmed by static review of the PR diff.

This is an automated hermes-sweeper review.

[kshitijk4poor]

Salvaged via #76647 with your commit cherry-picked and authorship preserved — thank you! Your fix is exactly right: the hard-cancel Event, the daemon-isolated protected provider call, and the full cooldown/lease rollback all survived review intact (174 tests, mutation-checked, E2E-verified).

The salvage rebases onto current main (your branch had drifted 51 commits and picked up some conflict-resolution damage in tui_gateway/methods_session.py), and adds one follow-up commit widening the same hard-stop semantics to _abandon_timed_out_gateway_turn, a timeout surface that landed on main after your branch point.

#76647 is armed to merge on green CI. Closing this one in its favor.