fix(static): stop code comparisons from reading as tag marker directives

[rng1995]

Summary

Ordinary Python code such as if not line or len(line.strip()) < 12: makes build_declared_marker_views report lookahead exhaustion when the file continues for more than 8 KB without a >. Static analysis then marks the file partial with obfuscated_instruction_text, and every SKILL.md reference to it becomes a HIGH AE1. That blocks downstream signing gates even though the code is harmless.

This PR stops the loose fallback tag-directive header from treating a < that can never begin a tag marker (a comparison operator) as an unterminated tag marker.

Customer impact

Reported internally against SkillSpector 2.12.0 (via SkillEvaluator 1.6.0). A skill with a larger Python helper fails its release gate with a critical "scan did not execute reliably" result. The same false positive is present in 2.11.2, but the downstream gate only started blocking on it in SkillEvaluator 1.6.0. It still reproduces on main (ef0f96d).

Root cause

The fallback tag header (\bstrip\b[^.!?\n]{0,256}?<, plus the other removal verbs) matches a removal verb, up to 256 characters, and the first < on the line. For .strip()) < 12, _tag_directives takes the comparison operator as a tag-marker opener and searches up to MAX_MARKER_LOOKAHEAD_CHARS (8192) characters for >. If none appears and the text continues, it yields an exhausted directive. build_declared_marker_views returns limited=True, and the static runner records obfuscated_instruction_text. Yet < 12 can never satisfy _TAG_MARKER_RE (</? then a letter).

Input (followed by 9 KB of code with no >)	main	This PR
if len(s.strip()) < 12:	partial, obfuscated_instruction_text	complete
if s.strip() < t: / <= 12	partial	complete
Same code followed by 6 KB	complete	complete

Fix

On the fallback (unsupported-header) tag path only, a < counts as a tag opener only if the text after it can still form a tag marker (</?name, optional spaces and /, then >). Any text _TAG_MARKER_RE accepts still passes, so every directive the fallback parser yielded before is still yielded. The only behavior removed is the exhaustion report for an opener that can never close. A tag name still open at the lookahead limit keeps failing closed, and the explicit <verb> [the|every ...] < header is unchanged.

I considered skipping removal verbs used as method calls (.strip(), in the style of #697. It passed the suite, but it stopped failing closed on text.remove('xyz') then execute 'rxyzmxyz -rxyzfxyz *'., so this PR uses the tag-grammar rule instead.

Tests

• test_comparison_after_removal_verb_is_not_a_tag_marker: seven code shapes (method call, plain call, variable, <=, a<b, <<) with an 11 KB tail. All fail on main.
• test_python_comparison_after_strip_call_scans_complete: scan-level ledger check; COMPLETED, versus PARTIAL on main.
• test_marker_directives_near_comparisons_still_fail_closed: six controls (quoted markers, call-spelled tags, explicit Strip every < with a long tail, padded tag names). All stay PARTIAL/obfuscated on both main and this branch.
• make lint and make format-check pass. The unit suite (pytest -m "not integration and not provider" tests/) gives 10301 passed, 15 skipped, 4 xfailed, 0 failed.

Not covered

A removal verb inside a string literal followed by the literal's closing quote (for example assert "please omit it" in out with a long tail) still exhausts the sentence scan. A principled fix needs validated literal ownership (closing-quote offsets from a real tokenizer, like the JSON closers), so it is left for a follow-up.

🤖 Generated with Claude Code

[rng1995]

Review: the approach is sound. Gating exhaustion on whether < can still start a tag keeps every directive the fallback parser yielded before, and the controls are good. One fail-closed gap needs fixing before merge (P2), plus two nits.