Ticket graphs

Tighten the `--engine cerebras` (GLM 4.7) worker path in ollama_agent.py so a less-steerable model can complete forge-rewrite pairs. Two live canaries (#414, #416) on 2026-07-15 failed the pair gate despite GLM producing a genuinely good 504-line rewrite, because of OPERATIONAL drift. Observed GLM 4.7 failure modes (from the canary logs): 1. IGNORED the embedded source. The ticket already carries the original bytes between ===ORIGINAL SOURCE=== / ===END SOURCE===, but GLM went filesystem-hunting instead (5x redundant Reads of docs/PLAYBOOK.md, then Globs). 2. WANDERED cross-module. It spent ~18 steps reading an UNRELATED module (forge2/audio_tagger/tagger.py) — unpredictable token burn. 3. Needed ~33 steps to first Write; the default --ollama-max-steps 12 cut it off mid-orientation. 4. Wrote to a SPURIOUS double-nested path forge2/<repo>/<repo>/mod.py instead of the flat standard forge2/<repo>/mod.py (forge_package.py's contract), so pair_loop's deliverable existence-check failed it. 5. Malformed tool-call schemas: Bash "command" passed as a LIST, broken shell quoting. DELIVERABLE (checkouts/current/congruency/tools/ollama_agent.py): when the engine is cerebras (worker role, openai flavor), prepend a tightened worker directive to the birth prompt AND adjust defaults: (a) "The ORIGINAL source you must rewrite is EMBEDDED in your ticket between ===ORIGINAL SOURCE=== and ===END SOURCE===. Base your rewrite on THOSE bytes. Do NOT search the filesystem for the original — it is not there (this is a fresh forge2 tree)." (b) "Write your COMPLETE deliverable to EXACTLY the given deliverable path, verbatim. Do not add package nesting, do not rename, do not create sibling files." (c) "Do NOT read, grep, or glob any module other than your own deliverable. You do not need to explore the repo." (d) Raise the default --ollama-max-steps for the cerebras engine to 40 (keep 12 for others). (e) Reinforce the Bash tool contract in the tool description the openai channel sends: command is a single STRING (argv0 must be an allowlisted program), never a list. ACCEPTANCE: a `--engine cerebras --ticket N` pair on a forge unit writes the correct FLAT deliverable within the step budget and passes the forge harvest gate (parse + pyproject + package-check), with zero reads of unrelated modules in its step log. Keep Claude as the fallback engine; this is about making GLM 4.7 carry routine forge rewrites to save budget (see the inference-routing-policy: GLM default, Claude fallback).

— browse congruency source (forge-parked tree)