skill: triage a run from its evidence pack

[lt-zeeshan]

What

Teaches the kane-cli skill to triage a run from its evidence pack, beyond the quick debugging flow.

• New references/triage.md: the triage steps, the window of steps to read, a screen table that sets each screenshot against what the agent meant to do, a six-part report, and four causes (the app under test, the test or objective, the environment, kane-cli).
• references/debug.md: adds step identity and the join rules (action_id is the join key, step numbers are not), where the intent of a step lives, and what a step folder holds. The debugging flow now says a step's screenshot is taken before its action runs, so the effect is in the next step's screenshot.
• New scripts/evidence-map.mjs: prints a step sheet for a pack. It joins every step folder to the actions log, the authored steps and result.yaml, and lists the error lines of the run logs. Node, no dependencies, reads the .evidence zip itself, so it works on macOS, Linux and Windows without unzip.
• Screenshot paths in the skill and the Kiro steering now read screenshot.*. Packs from recent builds hold screenshot.jpg, so the documented unzip … screenshot.png command found nothing on them.
• Kiro power: mirrored by hand, without the script. Mapping doc updated.

Checks

• cd skill-installer && npm test: 102 pass, 0 fail, 4 skipped (OS-specific). 29 tests are new, in test/evidence-map.test.mjs, all on hand-built packs.
• The three skill copies are identical.
• npm pack --dry-run ships skills/scripts/evidence-map.mjs and skills/references/triage.md, and no tests. The workflow's package check now looks for the script.
• The script was run on two real web packs (one authoring run, one replay).

Not covered

• Not checked against a mobile pack. The references describe only files seen in web packs.
• docs/user-guide/evidence.md still says screenshot.png. Left for a separate change.

🤖 Generated with Claude Code