v1.4.0 — auggie CLI channel (auto-preferred) + reasoning-truncation fix

Migration note: bare channel: cli now means cli_kind: auto — when the auggie binary is on PATH it is preferred over codex (one-time stderr banner). In auto mode the auggie try does NOT inherit cli_extra and takes its model only from auggie_model. Opt-out / revert: set cli_kind: codex on the member for exact pre-1.4.0 behavior. auggie is a billed channel (upstream API price +40% via your Augment plan) — dry-run counts it as billed.

Added

• CH2 second CLI kind: auggie (cli_kind: auggie|codex|auto, default auto → auggie preferred when detected). One auggie account serves all committee families (GPT/Gemini/Claude/Kimi/GLM). Hardening: prompt via --instruction-file (not argv), empty --workspace-root sandbox (no codebase-context injection into blind review), --output-format json envelope, --max-turns 1, --dont-save-session, --retry-timeout (measured: concurrent Augment 503 retries hung >7 min without it).
• CLI-path JSON repair round: cli seats (codex and auggie) get one self-repair call on unparseable output before falling to the next channel, matching the api path.
• channel_used records the concrete kind (cli:auggie / cli:codex).
• Default committee: seats A (gpt5.6-sol) and C (gemini-3.1-pro-preview) move to auggie with codex/api fallbacks; 4th-family alternative switches from grok (separate x-ai key, mostly list-only 404) to auggie kimi-k2.7.

Fixed

• Reasoning-model truncation on the api path (the OpenRouter gemini empty-shell bug): reasoning models can burn the whole max_tokens budget on reasoning (finish_reason=length, empty content); the old retry re-sent the identical request. call_model now doubles the budget per retry (cap 16000) and returns partial content best-effort on the final attempt.

Tests

• 137 → 162; real E2E: two auggie seats (explicit + auto-detected) generate 2/2 → stats consumed normally.