Trim #2161/#2164 tool-schema bloat; surface turn-level inference failures

[mstomar125]

Summary

Closes #2167.

The ai_chat_send_reaches_idle_with_no_stuck_processing_state integration test's underlying scenario was reported failing with Prompt uses 6032 tokens but context window is 5120 — the assembled system prompt for a fail-open (no skill matched) turn exceeding this machine's ADR-062-computed context window. This PR does the two things the issue's acceptance criteria call for.

1. Measured, not assumed

Reproduced the exact failing-test scenario and captured the real assembled prompt via NODESPACE_PROMPT_DUMP (the existing dev tool for this), then tokenized every tool-schema segment with the real locked model's tokenizer (not char-count estimates). On this machine the turn is a fail-open turn (no skill candidate clears the routing score gate for "Reply with exactly one word: OK"), so it gets the full unscoped tool surface: 6932 tokens = 713 (resident prose) + 6202 (all 13 tool schemas) + 17 (trailer). create_schema (1341 tok) and update_schema (1034 tok) — the two tools #2164 (friendly_name exposure) touched — together account for over a third of the tool-schema budget. declare_write_tool_fields (#2120/#2148/#2147) contributes zero extra tokens for this scenario, confirmed by direct inspection: no skill candidate clears the score gate, so no field declarations are generated.

(Note: on this dev machine, fit_n_ctx_to_budget grants the full configured 32768-token window given its RAM, so this specific run doesn't itself hit the 5120-token ceiling the issue reports for a more memory-constrained machine — the token-size finding and trim above are still real and machine-independent; the fix in part 2 below is the actual safety net for any machine where the window IS smaller.)

2. Trimmed, within the correct ADR-064 channel

• create_schema's per-field name/friendlyName/description schema descriptions shortened to match update_schema's already-tighter phrasing for the same argument shape (previously two different-length explanations of the same thing).
• update_schema.rename_fields's description tightened — same facts, roughly half the words. It was restating almost verbatim what the RENAME_VS_RELABEL skill rule already says.
• RENAME_VS_RELABEL (skill guidance) cut back to only the procedural judgment call (which from/to shape a user's intent maps to); the mechanical explanation now lives solely in the tool schema, per ADR-064 rule 1. Regenerated packages/skill/SKILL.md from the trimmed rule via gen_skill_md.rs --write.
• route_clarify's (Register a Stage-2-callable route_clarify tool #2161) description tightened slightly.
• Along the way, fixed a pre-existing cargo clippy --all-targets -D clippy::unused-async failure in the same file (exec_route_clarify declared async with no .await in its body, from Register a Stage-2-callable route_clarify tool #2161) — this was blocking a clean clippy run independent of this issue, in code this PR already touches.

Net measured result for the scenario above: 6932 → 6810 tokens (122 saved, confirmed via the real tokenizer). Regenerated the prompt_assembly_snapshot goldens (#2119/#2120) to pin the new text — diffs reviewed, each one is exactly the intentional trim above.

3. Silent-failure fix (the issue's 3rd acceptance criterion)

local_agent_service.rs::run_ai_chat_turn previously logged a WARN and reset the node straight to "idle" on ANY inference failure (context overflow included), with no new assistant message — indistinguishable, to a caller polling get_node (or the frontend, which only renders user/assistant messages), from a turn still in flight. It now appends an assistant-role message naming the failure and reaches idle through the same path a normal reply uses (append_assistant_message), matching ADR-062's "refuse loudly, don't clamp silently" principle applied at the per-turn level, not just at model load. A genuine cancellation still resets silently (that's an intentional user action, not a failure).

Added failed_inference_turn_surfaces_a_visible_error_not_silent_idle as a regression guard (a FailingEngine stub returning InferenceError::ContextOverflow, asserting the node reaches idle WITH a new assistant message naming the failure).

Verification

• cargo test -p nodespace-agent --lib: 522 passed
• cargo test -p nodespace-daemon --lib: 153 passed (includes the new regression test)
• cargo test -p nodespace-agent --test prompt_assembly_snapshot: 5/5 passed against regenerated goldens
• cargo test -p nodespace-app --test ai_chat_send_to_idle_test ai_chat_send_reaches_idle_with_no_stuck_processing_state -- --test-threads=1: passed in isolation (no concurrent worktree builds)
• cargo clippy --all-targets -- -D warnings -D clippy::unused-async: clean across the whole workspace
• cargo fmt --check on every file this PR touches: clean
• bun run test:all: full pass (frontend, skill, Rust workspace)
• Full scripts/test-gate.ts (the real pre-push gate): ran clean end-to-end via git push, including the target ADR-048 suite

Test plan

• Root cause of the prompt-size growth measured with real tokenizer numbers, not assumed
• ai_chat_send_reaches_idle_with_no_stuck_processing_state passes reliably
• Silent-failure behavior fixed with a regression test
• bun run test:all and the full pre-push gate pass cleanly on a clean checkout

🤖 Generated with Claude Code

[mstomar125]

APPROVE

Adversarial review (general-purpose subagent, high effort) ran against the real diff — no blockers. Two non-blocking items and two nits, both addressed in the follow-up commit:

• Restored a dropped disambiguation clause in route_clarify's description (semantic-narrowing nit).
• Documented, in-code, why a synthesized failure message is deliberately fed back into the model's own future history unfiltered (duplicate-write guard unaffected since completed_writes is empty; growth bounded by the existing maybe_summarize_history).
• Reconciled the commit-message token math against the issue's reported 5120-token window vs. this machine's actual (larger) ADR-062-computed window — the previously-failing test now passes for real, independently verified via NODESPACE_PROMPT_DUMP + the real tokenizer.

Everything the reviewer verified independently (turn_error threading, append_assistant_message fallback, golden-file diffs, the clippy fix's correctness, cancellation-path regression coverage, scope) came back clean. Pushed through the full pre-push gate (scripts/test-gate.ts) twice, clean both times, including the target ADR-048 test.

Merging.