v0.5.0

Ledger release: she watches tool outcomes directly now, catches repeated silent failures, and nags before work starts.

Features:

• PostToolUse ledger records edited files, verification runs, normalized failures, and redacted event details; Stop prefers ledger evidence and requires verification newer than the last edit.
• Silent-recovery guard injects context after repeated normalized failures, throttled per signature.
• Promise-no-act Stop precheck blocks final “I’ll do it next” style replies without a following tool call.
• Implementation prompts get a one-time evidence contract; question/explanation prompts do not.
• scripts/gate-stats.mjs summarizes block→pass conversion, gave_up rate, and gate counts from gate events.

Hygiene:

• Gate labels/summaries/block instructions moved into ko/en/ja corpora.
• State writes use tmp+rename atomic writes; hook entry scripts fail open with one-line messages.

Fixes from pre-release review:

• Nested hookSpecificOutput for context injection.
• Promise-no-act avoids ordinary sign-offs.
• TODOs in new untracked files are caught.
• Runner/mutating detection anchors to command position.
• Short/question prompts no longer manufacture requirement blocks.
• events.jsonl stores compact redacted summaries.
• Requirement evidence uses content-word majority matching.
• Stop reads transcript tail only.
• SessionStart greetings localized to ko/en/ja.
• Give-up confession is model-visible in final retry block.
• Ledger falls back to hardened failure-text detection without exit code.