Fix/operational recovery journal hardening

[gonos2k]

Harden operational recovery journals with tail repair, strict validation, and atomic writes

• Adds tail-repair logic to orphan process, worker recovery, and command recovery journals: on startup, any incomplete final append is either truncated or newline-terminated before new records are written.
• Strict JSON schema validation now rejects malformed or logically inconsistent complete records with a descriptive error (line number + reason) instead of silently skipping them.
• Introduces writeAllSync to guarantee full buffer writes by looping on short-write OS reports, used by all journal append and compact paths.
• Append paths now open with 'a+', capture original file size, fsync after write, and truncate back to the prior size on failure; compact paths use atomic temp-file rename with fsync.
• Adds automatic continuation of token-truncated assistant responses (up to 3 hidden prompts) in AgentSession, with a user-visible exhaustion notice when the cap is reached.
• Post-compaction maintenance is now time-bounded; failures surface as a warning message rather than stalling compaction. Scheduler suspension ownership is tracked through manual compaction and abort flows to prevent unintended turn starts.
• Risk: token estimation for non-ASCII and punctuation-heavy content is now higher than the previous chars/4 heuristic, which shifts compaction thresholds and cut-point selection.

^{Macroscope summarized fc9b6e4.}