fix: per-action write state so one write failure doesn't block the rest

[mattmezza]

Summary

Fixes #8 — a write failure (or skip/deny, or even a successful write) in manage_jobs blocked every subsequent write action in the same turn.

Root cause

_execute_tool tracked write actions with two request-wide values in request_state:

• write_executed (a single bool) — set True after the first successful write.
• write_decision (a single value) — set to the first write's approve/skip/deny outcome.

Because both are global to the turn, the guards at the top of _execute_tool rejected any later write once one write had completed ("Request already fulfilled") or been skipped/denied ("User skipped this action") — even a completely different action. Scheduling two reminders in one interaction, or retrying after a skip, failed.

Fix

Track write state per distinct action, keyed on a stable signature of the tool name + its arguments (_write_signature, order-insensitive):

• executed_writes — a set of completed action signatures. Only an identical repeat is suppressed, preserving the original guard against accidental double-sends.
• write_decisions — per-signature approval decisions, reused only for the identical action, so a skip/deny never leaks onto a different write.

Read-action approval caching (per match_key, with persisted ALWAYS rules) is unchanged.

Tests

Added regression tests in tests/test_tools.py:

• distinct writes both succeed after one completes (the issue's main case)
• an identical repeated write is still deduplicated
• skipping one write does not block a different one
• signature distinguishes distinct params and is key-order insensitive

Full suite: 310 passed. ruff check / ruff format clean.