[hardy30894]

Stacked on #11 (the eval harness); review that first.

Why

On 1,491 recorded Claude Code sessions, how often the model needs an offloaded block back depends mostly on what produced it and how big it is:

class	under 500 tokens	500 to 2k	2k to 8k
Bash:grep |	15%	38%	66%
Bash:sed	15%	48%	78%
Bash:cat	26%	52%	87%
Read		65%	79%

Every reopen costs an extra model turn, so offloading a large cat buys a turn and saves little.

What

• A reopen policy (crates/optimize/src/policy.rs): Laplace-smoothed rates per command class and size band, falling back to the class, then to a prior. A block may be offloaded only when its predicted rate is under the threshold, 0.4. Otherwise the optimizer runs with offloading off, so only inline codecs, which never need a reopen, apply.
• MCP optimize takes optional tool and command; exec uses its own command. Both classify with the same function eval uses.
• secondwind eval --learn fits the rates on the even-numbered sessions, replays the odd ones with and without the gate, prints both, and writes ~/.secondwind/policy.json. eval --gate replays with the active policy. The built-in table holds the classes with at least 300 offloads in the corpus above.
• The Claude Code mod passes the command with each block and compresses from 512 bytes instead of 6 KB.
• The query description no longer claims it helps text output.

Results on held-out sessions (learned on the other half)

	context saved	reopen rate	reopens per session	unfaithful
no gate	24.8%	42.8%	23.6	0
learned gate	10.5%	22.1%	4.8	0

For the mod, Bash only: the old 6 KB rule saved 6.6% of tool output with 5.9 reopens per session; the gate at 512 bytes saves 9.7% with 4.8, and cuts billed cost by about 21% versus 5%.

Tests

Policy learning, backoff, round trip and refusal of a bad file; the gate in evaluate; end to end through the binary: a large cat stays in view over MCP and exec while the same block unnamed offloads, a learned policy replaces the built-in one, and eval --learn writes a policy and reports held-out sessions. Mod tests: the command reaches the binary, output under 512 bytes never does. cargo test --workspace: 363 passed. Clippy clean.