Editing a symlinked SKILL.md never refreshes its imported shelf entry

Read #1410 alongside internal/skills/skills.go:277 and internal/resident/skills.go:226, then trace ReconcileImportedSkills at :715. Add the linked-file case to an internal/resident test and run the relevant tests. Done means the chosen safe policy is consistent for reading and digesting, nonregular targets cannot block discovery, use-skill.md is aligned, and invalidates records the behavior.

Found on santos/dev2 at 008363c98 (#1410). It reaches dev when #1410 merges.

What happened

A skill folder containing a SKILL.md symlink to a regular file loads through the link. Its import digest omits that link, so changing the target leaves the active shelf fact unchanged at the next launch. The manual says an edited SKILL.md is read again. Linked folders are already handled at this head and are outside this issue.

Replication

Deterministic (no model). In an internal/resident test, create a valid skill file outside a skill folder, link the folder's SKILL.md to it, call ReconcileImportedSkills, edit the target text, and call reconciliation again. Compare the active fact's digest and description. Today they remain the first version. A regular-file control refreshes.

Field (real models). On a fresh home, attach that skill in codeaf chat, edit its target, relaunch, and ask a question that distinguishes the two bodies. Allow OPENROUTER_API_KEY, deepseek/deepseek-v4-flash, about a minute and a small charge.

Where

internal/skills/skills.go:277, readSkillFile, follows the link. internal/resident/skills.go:226, contentDigest, skips its nonregular directory entry; ReconcileImportedSkills compares that digest at :715. The contrary manual promise is in internal/manual/chat/use-skill.md.

The fix

Choose one link policy together with the launch-hang fix being made on #1410 (a committed SKILL.md or .claude/settings.json that links to /dev/tty or a FIFO blocks launch): either safely follow regular targets in both reading and digesting, or reject linked files in both. This issue covers the digest half of that policy.

Acceptance

• e2e: relaunching codeaf chat after a target edit uses the new skill body, or consistently reports that the linked file is unsupported.
• Unit: a linked file and a regular file follow the same chosen policy, while a link to a nonregular target cannot block discovery.
• Align use-skill.md and record the changed behavior in invalidates.