test(learning): diagnose intermittent learned-skill publication failure

Problem

TestUsableAutoSkillsStockBuildPolicyMatrix/untrusted_project_ignored intermittently fails with only:

server: internal error: learned skill operation failed

The immediate retry passed without code changes, so this is currently an observability/reliability investigation rather than evidence of a deterministic policy regression.

Why diagnosis is insufficient

The test polls learned skills while asynchronous recovery/publication is active. ListLearnedSkills publishes before listing, but skillServiceError intentionally collapses publication and repository errors to the generic RPC error, leaving CI with no actionable underlying cause.

Relevant code:

• internal/app/usable_auto_skills_matrix_e2e_test.go:128-145
• internal/app/attempt_recovery.go:24-67
• internal/adapter/server/learnedskills.go:70-88,323-333
• internal/adapter/skillstore/store.go:199-245,431-527

A plausible transient source is the five-second flock-bounded skill-store path during publication/recovery, but the existing failure message cannot establish that.

Acceptance criteria

• On this matrix test's failure path, retain safe diagnostic context for the underlying publication/repository error.
• Do not expose skill content, credentials, or sensitive filesystem data.
• Keep the production RPC client error fail-closed and appropriately generic.
• Use the evidence to either eliminate a confirmed synchronization flaw or document and stabilize the intended transient behavior.

Observed in PR #957, workflow run 33855617933; it passed on immediate retry.