ATperson autonomous scheduler

The daemon ran only the perception half of the agent loop. Issue #140 adds the scheduling half: after each bounded perception cycle, the runtime can now decide, propose and execute — through the same gates, never around them.

Changes

• src/app/scheduler/cycle — one bounded cycle: recent committed ledger payloads become decision contexts (the ledger stays the sole observation authority), the guarded C23 decision layer accepts or abstains, and accepted plans are frozen as atperson-outbound-action v1 proposal documents under scheduler/proposals/<digest>.json. Abstention is a first-class counted outcome.
• Shared attempt atom — publish's gate-load → execute → budget → audit → journal sequence is extracted into outbound/attempt. The publish CLI and the scheduler now call literally the same code, so there is still exactly one write path.
• Approved execution — proposals execute only when the operator has approved their exact digest (atperson control approve), oldest first, gates reloaded from disk per attempt. A pause or revocation between cycles always wins. Executed proposals are consumed; refused ones stay for the operator.
• Daemon wiring — the scheduler runs in the daemon's on_cycle hook, off by default (ATPERSON_SCHEDULER=1), with per-cycle bounds on contexts, proposals, executions and wall clock.
• autonomy status — now reports pending proposal count alongside phase and run detail.

Safety

• Proposal files are written atomically (temp + rename); an interrupted cycle leaves no partial state.
• Execution remains the atomic idempotent unit it was: frozen rkey, putRecord retry semantics.
• No new write kinds, no C23 decision changes, no gate weakening. The scheduler composes.

Verification

• New offline atperson-scheduler test: fake writer, injected clock, real ledger and gate files. Six scenarios including disabled inertness, unapproved freeze, approved execution, pause refusal, no-rewrite and abstention.
• Core-only build 53/53, network build 59/59, no new warnings.
• PR #145.