skillscript-runtime v0.38.0

Upgrade impact: none (additive). Four new pipe filters, a lint false-positive removed, and a discovery surface widened — no behavior change to existing skills.

Pipe filters — line-slice family + pluck (Perry build spec; each cleared the project's "a real authored skill demonstrated the gap" bar):

• head:"N" / tail:"N" / lines:"M-N" — line-slice a string (one shared impl). Split on \n, CRLF-tolerant, and a terminal newline's trailing empty line is dropped (so tail:"1" of "a\nb\n" is "b", not ""); lines:"M-N" is a 1-indexed inclusive range. Never throws — a bad / negative / out-of-range count clamps to what exists or yields "". Security payoff: file_read(spool) -> L then ${L|tail:"N"} tails a log fs-read-only, with no tail/grep shell-binary allowlist grant and no unsafe=true; a file_read help cross-reference now points authors to it.
• pluck:"field" — project one field from each element of an array of objects → the array of field values (emitted as a JSON-array string, so it composes with in / not in / |length / |contains, all JSON-string-of-array tolerant). Omits an element whose field is absent/null or that isn't an object; single-level field only. Turns a hand-rolled dedup-by-id loop into a clean projection. Current idiom binds the projection then membership-tests ($set IDS = "${SEEN|pluck:"id"}" then if ${M.id} not in ${IDS}:); an inline filter chain on the in RHS is a pending grammar extension.
• Both operate on the stringified input (the filter layer is type-blind); pluck additionally content-parses and throws on non-array input.

approved= no longer flags as unknown-connector-arg. The mutation-gate auth kwarg is popped by the runtime before the connector sees args, so validating it against the tool's input schema was a false positive that directly contradicted unconfirmed-mutation's "add approved=" remediation. timeout= and approved= are now a named reserved-dispatch-kwarg set, exempt from arg validation.

observed_output_shape now rides runtime_capabilities({tool}). The last-observed return shape was surfaced only in skill_preflight (which needs a skill already referencing the tool); it now also attaches to the tool fetch authors reach for to answer "what does this return?", and the tool description advertises it.