Telemetry: distinct command outcomes, per-kind failure streaks, periodic flush, per-tool output schemas

[cf-pages]

Targeted: 0.5.0 (reliability program). Baseline: release/0.3.0.

Four measurement gaps that keep real failures invisible:

1. Non-zero exits count as success. processes.py:280 hardcodes "ok": True for any command that ran; a failing test suite is a successful tool call and that is correct MCP semantics — the bug is that telemetry collapses both into one ok. Add an outcome dimension: exited_0 | exited_nonzero | timeout | signal | spawn_error, bucketed in tool_summary/session_end.
2. consecutive_failures is a single global slot (telemetry.py:400-407): any success of any tool resets it, so alternating failures never accumulate (17 consecutive apply_patch failures interleaved with 6 read_file successes peaked at 5 in the 0.2.x sample). Track per (tool, error_code), cleared only by a success of the same tool; bounded (LRU ~16).
3. Summaries only flush at close. 0.3.0 made tool_error immediate and documented that long-lived HTTP processes rarely emit session_end; add periodic incremental flush (every N calls or T minutes).
4. One generic output schema for the whole catalog (tool_output_schema()): command_id, next_action, truncated, status, exit_code are undeclared. Publish per-tool output schemas (also a prerequisite for the apply_changes program).

Privacy red lines unchanged from 0.3.0: closed schema, no paths/content, allow-listed clientInfo labels.