v2.2.0: fewer tokens in each agent session

Added

• ContextOpts.continuity and PromptHookOpts.continuity ('full' | 'pointer', default 'full'). Library callers opt in to the one-line form. The remote hook route keeps the full text unless its caller opts in.

Changed

• The prompt hook and hippo context print another session's task snapshot and handoff as one line. The line names the session and the command that prints the full text (hippo session latest --id=<id> or hippo handoff latest --session=<id>). A session id that a shell or the flag parser could misread prints the full text instead. Your own session, a shared store, --format json, MCP hippo_context and the HTTP routes still print the full text. In the Z0 round 2 eval, a finished session's handoff was 2,407 of the 3,708 characters the hook added to each new session.
• The Claude Code and Codex instruction blocks no longer ask for hippo context --auto on every task. The prompt hook already injects pinned rules and matching memories, so the blocks ask for the call only when no hippo block appeared in the session. hippo init refreshes blocks it installed; hand-edited blocks stay as they are. The Claude Code plugin (0.6.2) drops its SessionStart context call for the same reason, and with it the session-start hint to run hippo init when no store exists. A hooked session no longer gets the git-aware recall at session start; the recall matched to each prompt (pinnedInject.promptRecall, on by default) still runs. The per-prompt hook does not record what it shows, so those memories are not strengthened, and hippo outcome with no --id applies to the last recall that was recorded.
• hippo no longer prints Node's SQLite ExperimentalWarning on each call. An agent read that warning on every hippo command. Other warnings still print.
• ContextCost.snapshot and ContextCost.handoff take a second mode argument. Implementations written for the old signature still type-check, but a caller that invokes them must pass the mode.