[hardy30894]

Bumps every crate, the npm package with its five platform packages, and the Python package to 0.5.0, and the Claude Code mod to 0.2.0 so marketplace installs pick it up with claude plugin update. Same nine release files as v0.4.0, plus the mod's plugin.json.

Nothing is published by merging. Publishing happens when the v0.5.0 tag is pushed on the merge commit. The npm token in NPM_TOKEN expires October 29, so tag before then or the npm leg fails.

Verified: every edited manifest parses, the release build reports secondwind 0.5.0, the mod validates and its 8 tests pass. Main's CI already passed on all three platforms for the code in this release; this PR changes versions and notes only.

It is a minor bump because setup --hook no longer installs anything and figures read from recorded sessions change.

Release notes

Upgrade notes

• The Bash hook is retired. Claude Code checks permission rules against the command a
  PreToolUse hook returns, and the hook rewrote every Bash command to secondwind exec -c '...'
  while answering "allow", so Bash deny rules stopped matching and prompts were skipped. secondwind hook now passes every command through unchanged, setup --hook installs nothing, plain
  secondwind setup removes a hook an earlier version installed, and secondwind check warns while
  one is still there. Use the Claude Code mod instead (below).
• Offloading depends on where a block came from. When a caller names the tool and command (MCP
  optimize takes optional tool and command; exec uses its own), output that is usually read
  in full, such as a large cat or Read, stays in view and only inline codecs apply. Callers that
  name nothing behave as before.
• Figures read from recorded sessions drop. The trace reader counted images and PDFs as base64
  text and wrapped every tool result in JSON the model never saw. On the author's sessions that
  made tool output look seven times larger than it is, so cache-savings and the audit report less.

Security

• The retired Bash hook let Bash commands run without the user's deny rules or prompts (above).

Added

• secondwind eval replays your recorded sessions through the optimizer and reports, per tool and
  command, what was offloaded, how often the model later needed it, net tokens, and whether every
  block came back intact. --learn fits reopen rates on half your sessions, replays the other half
  with and without the gate, and writes ~/.secondwind/policy.json; --gate and --query replay
  with the gate or with your latest message as the query; --blocks writes one line per result.
• A reopen policy per command class and size decides whether a block may be offloaded. The
  built-in table comes from 1,491 real sessions; a learned policy.json replaces it. On held-out
  sessions the learned gate cut reopens per session from 23.6 to 4.8, and no block of 138,209 came
  back different from the original.
• Claude Code mod 0.2.0: sends the command with each block, so a large cat stays in view, and
  compresses from 512 bytes instead of 6 KB. For Bash on held-out sessions it saves 9.7% of all
  tool output instead of 6.6%, with fewer extra turns. Tests now pin the pane's off, empty, win and
  loss states.

Fixed

• optimize's query description no longer claims it helps text; it applies to JSON record
  arrays.