fix: protected-span tell eligibility — procedural text no longer creates evidence obligations

[Nazim22]

Closes the phantom-obligation defect found by running Leadline on a real session.

Bug. Tells matched anywhere in a prompt, including inside quoted text, code spans and shell commands. A runbook-style instruction containing git diff abc..def produced a repository obligation nobody asked for, and every later unrelated tool call was then graded against it. In one real session that single false obligation generated 32 downstream corrections.

Fix. A lexical scanner marks protected spans (quoted text, inline code, fenced code, parenthesised shell groups containing ;/&&/||), decomposition no longer splits on delimiters inside them, and tells overlapping a protected span are ignored. The naked git diff tell is removed — a direct command already prescribes its own source, so routing adds nothing there.

Deliberately not a general speech-act classifier: that would become another hand-authored phrase inventory one layer earlier. Genuine imperative evidence requests are unaffected.

Verification (reproduced independently by the reviewer):

• npm test 237/237 in a clean worktree
• Benchmark unchanged: 62.5% first-route accuracy, N=16 routed
• Satisfaction simulation 7/7, gate-gaming bypasses 0/4
• The original production prompt now yields contract.steps = 0 — verified by replaying the exact cron text that caused the 32 phantom decisions
• Positive controls behave identically to base (checked against 8d2b7a7 — no routing regression)

Author: Daedalus. Reviewed against exact git objects (commit e355a5f, tree fa90841, patch-id b7925f7).

🤖 Generated with Claude Code