feat(libsy): LLM-driven tool-signal discovery with warm-up gate

[sabhatinas]

What this does

stage_router currently decides routing from a fixed regex/name-table extractor (tool_signals.rs). This adds a second option: an LLM judge that watches each turn's tool activity and discovers its own severity/category signals instead of matching hardcoded patterns — useful for agent harnesses the regex tables weren't tuned for.

Both extractors produce the exact same ToolSignals shape, so the router's actual picking logic doesn't change at all — you just choose which one fills it in.

Warm-up

Since the judge builds its understanding turn by turn, a new warmup_turns setting holds routing on the default (efficient) tier for a session's first N turns while the judge still runs and calibrates in the background. Once warm-up elapses, routing decisions kick in for real.

The judge's discovered vocabulary is also shared across every task in a benchmark run (not reset per task), so it keeps getting better calibrated the more it sees.

How to turn it on

Two new TOML fields on a stage_router route:

signal_discovery_judge_target = "opus"   # judge model; omit to keep the old static extractor
warmup_turns = 5                         # hold the default tier for the first 5 turns

Status

This is an experimental prototype — validated so far on a handful of TB 2.1 tasks across Claude Code and Codex, not yet run at benchmark scale. Opening as a draft while that validation continues.