fix(scheduler): use explicit execution prefix for scheduled task injection (#2073)

[bug-ops]

Summary

Fixes issue #2073: Scheduler task injection format is ambiguous — LLM cancels bash tasks instead of executing them.

Problem

When the scheduler fires a deferred task via the custom_task_tx injection path, the injected message was formatted as:

[Scheduled task] bash -c 'echo hello > /tmp/output.txt'

This is ambiguous. With gpt-4o-mini, the LLM interprets this as informational metadata and calls cancel_task instead of executing the bash command.

Solution

Replace the ambiguous [Scheduled task] prefix with the explicit execution directive:

Execute the following scheduled task now: bash -c 'echo hello > /tmp/output.txt'

This imperative wording prevents LLMs from misinterpreting the annotation as metadata.

Changes

1. Agent loop (crates/zeph-core/src/agent/mod.rs): Introduced SCHEDULED_TASK_PREFIX constant and updated the injection format at line 2101.
2. Scheduler fallback (crates/zeph-scheduler/src/scheduler.rs): Updated oneshot fallback format at line 316-319.
3. CustomTaskHandler (crates/zeph-scheduler/src/handlers.rs): Updated handler fallback format at line 33 + updated related test.
4. Documentation: Updated all references in SKILL.md, scheduler.md, daemon.md.
5. Unit test: Added test_scheduled_task_injection_format to verify the new format (references the constant to prevent silent drift).

Verification

• ✅ All pre-commit checks pass: cargo +nightly fmt, cargo clippy, cargo nextest
• ✅ 6265 tests pass (+1 new unit test)
• ✅ All validators approved: tester, security, impl-critic, reviewer
• ✅ No unrelated changes bundled

Test Plan

1. Unit test: test_scheduled_task_injection_format validates the new format string.
2. Integration: Existing scheduler injection tests continue to pass.
3. Live test: Next CI cycle will verify with gpt-4o-mini that bash tasks are no longer cancelled.

Issue Resolution

Resolves #2073 completely. The format change applies to all three scheduler injection points (agent loop, oneshot fallback, handler fallback), preventing LLM misinterpretation across all code paths.