fix: don't trust a partial dev-commands.json cache as fully discovered

Fixed

• Command discovery (discover-commands.sh) no longer trusts a partial .coding-crew/dev-commands.json as fully discovered. It used to skip re-running discovery the moment the cache file existed at all, regardless of which of the six fields (test/lint/typecheck/install/env/credential_target) it actually contained. That let a partial write — e.g. ensure-deps.sh's own install_mode/docker_service fields landing first because an earlier sprint's model dispatch failed (expired credentials, a timeout) before ever writing the six discovery fields — permanently starve every later sprint of full discovery, since the file "already existed." The skip now requires all six fields to be present (any value, including JSON null, which still means "checked, confirmed no such command") before trusting the cache; a cache missing even one field falls through and re-runs discovery.

feat: prune skills dropped from a --skills list

Fixed

• install.sh now re-execs into a newer bash instead of dying with declare: -A: invalid option on macOS. Apple's stock /bin/bash is 3.2 (frozen there since the GPLv2→GPLv3
  switch) and install.sh's own #!/bin/bash shebang is resolved by the kernel from that
  hardcoded path, not from $PATH — so a newer Homebrew bash already on a user's PATH was
  never picked up just by running the script (which is what bootstrap.sh's exec "$INSTALL"
  does). The registry-read cache needs declare -A (bash ≥ 4). The script now detects bash < 4,
  hops to a findable newer bash (Homebrew's common install locations, or PATH), and exits with
  brew install bash guidance if none exists.

Added

• install.sh claude --skills a,b,c now treats that list as the full desired skill set: a
  name present in a prior --skills install but missing from this one is uninstalled, instead
  of being silently left on disk forever. write_manifest()'s merge only adds/updates keys, so
  without this a shrunk --skills list had no way to actually drop anything. The new
  prune_skills_not_in() shells out to uninstall.sh --skill — the existing single writer of
  skill removal — rather than duplicating that logic, then deletes the pruned key from
  manifest.json so the merge doesn't resurrect it.

chore: remove caveman and improve-codebase-architecture skills

Removed

• The caveman and improve-codebase-architecture skills are gone, along with their
  registry.json entries, install.sh/bootstrap.sh/unbootstrap.sh/uninstall.sh examples,
  doc references, and structural tests. Both were third-party skills (JuliusBrussee/crew-caveman,
  mattpocock/skills) this repo no longer maintains as part of its distribution.

fix: clear stale sidecar report.json before each dispatch

Fixed

• runWorker/runReview/runTriage now delete each dispatch's .report.json sidecar
  before writing that round's prompt, instead of leaving a prior round's (or a prior resumed
  sprint's) sidecar sitting at the same fixed path. waitForSidecarReport in dispatch.mjs
  treats existsSync as "found" the instant the herdr wait settles, so a stale sidecar from an
  earlier dispatch to the same slug was read back as this round's verdict whenever the coder's,
  reviewer's, or triage agent's turn died before writing a fresh one — a false verdict rather
  than the intended pane-scrape fallback.

feat: extend the sidecar-first herdr dispatch policy to review and triage

Changed

• The <slug>.report.json sidecar-first dispatch policy (already used for the crew-coder
  worker) now also covers the reviewer and triage agents. dispatchViaHerdr checks each
  role's report file before scraping the pane at all: the review and triage prompts now name
  a reportPath and instruct the agent to write its verdict there as its last action, falling
  back to the same fenced-json-in-prose reply only if that file never lands.
  parseReviewReport/parseTriageReport accept the parsed sidecar and prefer it over the
  captured text entirely, same policy parseWorkerReport already had. Closes the same class of
  herdr read-capture failure (empty/garbled pane render) for review and triage that the worker
  fix already closed — a genuine verdict no longer depends on a terminal scrape succeeding.

fix: report render line count and echo-found status on herdr empty-reply failures

Changed

• A herdr DISPATCH-FAIL outEmpty=true line now reports the captured render's line count and
  whether the echoed prompt was found anywhere in it (lines=N echoFound=yes|no). Previously
  the log's only diagnostic was the last 400 characters of the render — which is always just
  Claude's bottom status-bar chrome whether the reply capture failed or succeeded, so it couldn't
  distinguish "the read's --lines window (or the pane's alt-screen buffer) never captured the
  echoed prompt at all" from "the echo is there but extractHerdrReply's end-marker pattern
  didn't match." echoFound is computed by herdrEchoFound, sharing herdrIsEchoLine with
  extractHerdrReply itself rather than reimplementing the echo check.

feat: unify review/triage verdict reporting on a shared JSON contract

Changed

• The reviewer and triage agents now report their verdict as a fenced json block
  ({"branch", "slug", "verdict", "detail", "findings"} / {"fixable", "category", "detail"})
  instead of column-0 ## Branch:/AC:/Review:/FINDING: and FIXABLE:/CATEGORY:/DETAIL:
  lines, mirroring the structured-result contract crew-coder already used. The old shape had
  three independent hand-rolled parsers of the same aggregate review report file — a Node regex
  gating the merge, an awk in crew-summary.sh's rollup, and a second awk in
  promote-findings.sh's remind — and they drifted apart on how much whitespace a
  herdr-captured review header could carry before a line no longer matched. A genuinely
  successful, all-met retry review could land indented and without the ## after an earlier
  not_run stub in the same file and still be silently dropped from the summary and the findings
  count, even though the merge/AC gate (whose regex already tolerated leading whitespace) had
  parsed and acted on the real verdict correctly. JSON's whitespace-between-tokens-is-insignificant
  grammar closes the whole class rather than one more anchor at a time — crew-summary.sh and
  promote-findings.sh now both read the aggregate file through one shared parser
  (orchestrator/review-rollup.mjs, over report.mjs's new parseReviewAggregate) instead of
  reimplementing it twice. report.mjs's parseReviewReport/parseTriageReport prefer the json
  block and fall back to the old markdown shape, so an un-migrated reviewer/triage agent still
  works. mark-not-run's stub now emits the same json shape (verdict: "not_run").

fix: notify the triggering herdr pane even when a sprint fails before it starts

Added

• Under HERDR_ENV=1, crew-afk now pushes the sprint's outcome into the triggering pane
  ($HERDR_PANE_ID) once the run finishes, stalls, or errors, instead of leaving that pane to
  poll on a long interval for the same information — see notifyTriggeringPane's doc comment.

Fixed

• That nudge only fired once a sprint had actually been initialized, so a preflight or
  feature-slug resolution failure — the fastest way a run can end — left the triggering pane
  with nothing to stop polling for. main()'s run is now wrapped in a single try/finally
  from before preflight onward, so every exit path, including one that never got as far as
  creating a Sprint, still reaches the notify call.

fix: fail fast on a herdr pane blocked mid-turn, log last-seen agent_status on empty replies

Fixed

• A dialog appearing mid-turn on a herdr pane (after --wait settled on a stale idle/done
  match) polled uselessly until the dispatch's whole deadline expired before reporting an
  unexplained empty reply. Unlike idle/done, a blocked agent_status never resolves on its
  own — it stays blocked until a dialog is answered — so dispatchViaHerdr now fails immediately
  on that status (from both the direct check and from inside waitForHerdrIdle's own polling)
  with the pane's actual dialog text, the same way it already does for a dialog that's there
  before the prompt is even sent.
• A DISPATCH-FAIL outEmpty=true log line gave no way to tell a genuinely empty reply from
  one that simply ran out of deadline while the pane was still busy. The line now includes the
  last-seen agent_status alongside the rendered tail.

fix: check agent_status directly before giving up on a busy herdr pane that never errors on read

Fixed

• A herdr dispatch could still report an empty reply even after v1.29.79's agent_not_idle
  fix, when the pane was busy in a way that never made agent read itself error. The same
  --wait "does not track turns" gap could leave a read succeeding with code 0 while only
  catching the pane's live status footer (tool-call counters, a running timer) — indistinguishable
  from a genuinely blank reply, and the fixed-delay backoff that follows only covers a few
  seconds of flush lag, not minutes of real work. dispatchViaHerdr now checks agent_status
  directly once that backoff is exhausted: if the pane is still actively busy, it polls until
  idle/done (bounded by the dispatch's own deadline) and reads again, instead of giving up while
  the coder is still working.