[kesonm]

Summary

Fixes #6827 on rust-preview.

• Preserve overdue triggers when resuming with Misfire::Latest, including one-shot tasks. A trigger due exactly at resume time remains claimable; pending fires retain their identity.
• Select the next timer using the same cutoff used to load ready tasks, so a deadline crossed during a tick causes an immediate next pass instead of losing its wake.
• Reconcile scheduler errors with exponential retry delays from 1 to 30 seconds, without postponing earlier scheduled deadlines or replacing the original pending fire.

Verification

Tested on macOS 26.5.2, Rust 1.99.1, based on 7d572089b. The locally installed development build contains this patch.

• cargo test --locked -p maka-scheduler --lib: 8 passed. The new overdue-resume regressions also failed against the original policy source in an isolated harness.
• cargo test --locked -p maka-agent-native --lib scheduler::: 2 passed, covering deadline crossing and bounded retry wakes. Deadline crossing is a deterministic unit test, not a desktop timing reproduction.
• Built the producer_reopen test executable and ran its scheduler_ cases with runtime CARGO_BIN_EXE_maka pointing to the installed fixed binary: 3 passed, 0 failed, 54.55 seconds:

scheduler_resumes_overdue_latest_reminders_through_public_host ... ok
scheduler_retries_original_fire_when_workspace_returns_without_host_mutation ... ok
scheduler_user_approved_due_fire_fails_original_run_without_redispatch_after_reopen ... ok

The retry test temporarily moves an isolated workspace directory, observes an unexecuted pending fire, restores the same directory without a Host mutation, and verifies automatic admission with the original fire ID. It also exercises cold-reopen deduplication. Tests use temporary data and a loopback model fixture, not production credentials.

• just build: passed; installed app signature verification and desktop launch passed.
• cargo fmt -p maka-scheduler -p maka-agent-native -p maka-cli --check, git diff --check, and node scripts/ci.mjs check licenses: passed.
• Full just check and all reverse-dependent suites were not run locally; the focused tests above are the local evidence. A merge-tree check against rust-preview at d68c10f60 is conflict-free; the installed build was not rebased onto those later upstream changes.

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex assisted with implementation, regression tests, local build and verification, and this description. The commit includes its Generated-by trailer.

Checklist

• Tests cover the change and fail without it — original-source failure was verified for the resume regressions; a complete before/after run of every new test was not performed.
• Format, typecheck and the affected suites pass locally — build, formatting and the focused tests above passed; full affected-suite coverage is left to CI.

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No