Replace skill-failure noop with manual-analysis fallback

[theletterf]

Summary

Reverses #133's strict abort rule for the openings, applies-to, and style sweeps. Real-run evidence showed it firing prematurely and suppressing legitimate output.

From elastic/docs-content openings run 25378183549:

13:09:34  ✗ skill(docs-page-opening-optimizer)  Skill not found
13:09:37  ✗ skill(docs-page-opening-optimizer)  Skill not found
13:09:41  ● noop  "skill unavailable — confirmed after exact-form attempts"
13:14:36  ● create_issue  "shard 20/28 — 20 pages"   ← agent eventually wanted to file
                                                        but the safe-output system
                                                        suppressed it because noop
                                                        already fired

The agent's tool serialization keeps logging the skill call as skill(docs-X) (no skill: prefix), and that form always returns "Skill not found" from Copilot CLI. Whether the agent really sends that form or it's a log-rendering quirk doesn't matter — my hardening committed the agent to noop after two of these "failures", which then locked the safe-outputs system out of accepting the create_issue call the agent emitted 5 minutes later after manual analysis.

Fix

Lift the frontmatter sweep's softer pattern across the three strict-abort sweeps:

-1. Invoke with the exact form `skill(skill: docs-X)`.
-2. If "Skill not found", retry once more with the same exact form.
-3. Only after a second exact-form failure, abort by calling `noop`.
+1. Try invoking with the exact form `skill(skill: docs-X)`.
+2. If that fails or returns "Skill not found": fall back to manual analysis
+   using bash + judgment. Produce findings as you would have from the skill
+   output, note "skill X did not produce output; findings are agent-only"
+   in the issue body's Notes section.
+
+Only call noop if even manual analysis produces no high-confidence findings.

Each sweep gets a category-specific manual-analysis recipe:

• openings — bash to read first 10 lines, grep for generic-word H1s, check opening-paragraph length
• applies-to — scan frontmatter YAML, check applies_to against allowed values
• style — grep for banned phrasing, second-person inconsistencies, missing alt text

Sweeps unchanged

frontmatter already had this softer pattern (if one fails, report only the other and note the skill failure in Notes). coherence and staleness use different mechanisms. typos doesn't import any skill.

Test plan

• Merge + cut release moving v1.
• Re-dispatch docs-openings-sweep. Confirm: even if the skill returns "Skill not found", the agent produces a fix-issue with manual-analysis findings rather than noop'ing.
• Confirm the "skill X did not produce output" Notes line appears in the resulting issue body.

🤖 Generated with Claude Code