Harness reliability: run termination, context-safety margins, compaction fidelity

[anandgupta42]

Long-running non-interactive sessions exhibit several reliability failure modes:

1. No termination path — runs that finish their task keep consuming turns until wall-clock timeout; a majority of long-run wall time is burned after the work is done, and clean-exit rates sit near 15%.
2. Context-window overflow at the provider — the chars/4 token estimate undercounts dense SQL/JSON by up to ~1.5x, so requests sized near the declared window are rejected by the provider (400), and a single oversized tool result can jump a small conversation past the hard limit in one step.
3. Task drift under repeated compaction — iterated summarization loses the original task statement, producing re-grounding loops and zero-write sessions.
4. Embedder gaps — external drivers invoking run don't get run-mode semantics without env plumbing; explicit opt-outs were not consistently honored.

Tracking issue for the harness reliability changeset addressing these (run termination protocol, compaction fidelity + context safety margin, per-tool-result dispatch cap, run-mode defaults, accounting/attribution).