Spill, don't truncate: oversized tool outputs go to a session artifact, cap note cites the path

[justrach]

From the prime-agent analysis (PrimeIntellect-ai/prime-agent, 2026-08-06). Their model keeps every large result addressable (a Python object in the kernel namespace) and only caps what reaches the transcript (64KB/cell). The transferable idea needs no kernel: never destroy data at the cap point.

Today: capOversizedToolOutputs (#196) caps an oversized tool output per-output when over the context budget. The elided content is gone; the model re-runs the tool if it guessed the wrong slice.

Proposal: before capping, write the full output to a session artifact (e.g. .graff/sessions/<id>/artifacts/tool-<call>.txt) and make the elision marker actionable: [truncated: full output (N bytes) at <path>]. The model can then read/grep exactly the slice it needs.

Notes:

• Same call sites as fix(agent): per-tool-output cap + anthropic/openai overflow recovery (#193 follow-up) #196 (anthropic/openai/codex paths); same safety class as the cap itself (no WS-close bracketing needed, per the fix(agent): per-tool-output cap + anthropic/openai overflow recovery (#193 follow-up) #196 analysis).
• Bound artifact growth: per-session byte cap, artifacts deleted with the session.
• Only spill when a durable session exists; --no-session-style runs keep plain truncation.

[justrach]

Shipped in v0.0.241 (merge train PR #446; released 2026-08-06). Verified by golden evals, CI, and live gpt-5.6-sol runs — see the release notes and the live-eval findings referenced in #440/#445 for measured impact.