Returns the top-K most-similar pairs to `query`, each as a dict
    with pair_idx, session_id, premise_preview (≤200 chars),
    correction_preview (≤200 chars), and cosine score in [-1, 1].

    Requires `features.npz` (built by `weighted-compact bootstrap`)
    and the [baselines] extra (sentence-transformers). On the first
    call the e5-multilingual-small model is loaded into memory and
    cached for subsequent calls.

    On missing substrate or import error returns a single dict with
    an `error` key rather than raising — the stdio loop must not die
    on a misconfigured client.
    

    Hides the source pair (so it can be used as ground truth in
    reconstruction-QA) and returns the remaining session pairs
    ranked by the chosen scoring source, truncated to keep
    `(1 - k_drop)` of the session.

    Args:
        source_pair_idx: index into pairs.jsonl identifying the
            pair whose session should be compacted. The pair itself
            is hidden from the output.
        k_drop: fraction of session pairs to drop (0.0 = keep all,
            0.9 = aggressive compaction).
        ranker: scoring source. Currently only "importance" is
            supported here (query-aware rankers belong in
            search_pairs).
        rem_decay: if True, multiply scores by the nightly REM-decay
            map (requires `weighted-compact rem-pass` to have run).

    Returns a dict with `markdown` (the compacted context, ready to
    paste into a prompt) and `meta` (budget-transparency: pairs_total,
    pairs_kept, input_chars, output_chars, tokens_estimate,
    compaction_ratio, signals_top3, ranker label).

    On invalid input or missing substrate returns a dict with an
    `error` key rather than raising.
    

    Returns a dict with pair_count, session_count, has_importance,
    has_rem_decay, rem_decay_ref_iso (ISO timestamp of the last
    REM pass), and signals_present (names of the importance mixture
    components). Includes path of the substrate workdir.

    Use this on connect to know which other tools will succeed
    without paying their setup cost first. Always returns a dict
    even when nothing is built.