v3.9.0 — a re-attempt of a direct job blocked itself, forever

A partner tested 3.8.5 on a fresh repository before recording a tutorial (a small salon booking app) and it never
reached code: /v:dispatch demanded a taxonomy only /v:onboard creates, every re-attempt stopped on a "baseline
violation", Fable was spent under the cost-aware stance, and the plugin engaged only through explicit commands.
Three independent reviews (Claude, Codex, Gemini) traced each to code; two were reproduced in throwaway repos. This
release fixes all of them. Nothing here is measured yet: the end-to-end run with time and usage is the next step.

Fixed — a re-attempt of a direct job blocked itself, forever

resume-prepare removed the baseline pin but kept the first attempt's snapshot digests, so the gate charged the
job with its own new jobs/<id>.baseline on every attempt. Task 0 (shared_foundation) is always direct, so this
hit job 1 of every FULL run. Retiring an attempt now drops only the lines for the files registration rewrites
(jobs/<id>.baseline, jobs/<id>.test-contract.json, lane-map.json, preexisting/<id>.verified.txt); everything
else stays digest-bound, so a halted attempt's out-of-lane leftover is still a violation. A worktree job's snapshot
is deleted and re-provisioned. Behind that fix sat a second one: once the gate passed, the integration authority
refused every direct re-attempt as contradicted, because Record archives the previous result as
results/attempts/<id>.<n>.json and the gate's exemption list did not name it. The gate now lists that archive
before Record writes it. A superseded receipt (receipts/<id>.gate.superseded-*.json) is exempt by name.
tests/test-reattempt-direct.sh runs attempt 1, /v:resume, attempts 2 and 3 and finalize-wave.

Fixed — a hand re-run of /v:dispatch charged the job with every commit since attempt 1

Committing anything between attempts (the partner committed the taxonomy) blocked the direct job on that commit,
and the only warning went to the Workflow agent. /v:dispatch <run-id> now runs resume-prepare before emitting;
if that is skipped, emit and register-lane exit 2 naming /v:resume <run-id>. /v:resume re-emits the workflow
too, so re-attempt escalation and the current retry settings apply. A second registration inside the same attempt
no longer re-photographs a worktree (it used to forgive the worker's own out-of-lane writes).

Fixed — the taxonomy /v:dispatch requires was never created

/v:init Step 4g drafts .claude/compound-v-impact-taxonomy.yaml with compound-v-onboard.py draft-taxonomy, shows
it and commits it on a yes. Re-running /v:triage after the taxonomy appears mints a new record that supersedes the
old one instead of crashing on "already exists with DIFFERENT content"; an identical re-run reuses its record. The
null-digest refusal names the taxonomy, /v:init, /v:onboard and the re-triage, and
compound-v-preeval.py record-status says whether a record is dispatchable. --require-triage now also refuses a
manifest bound to a superseded record, or whose tier or flavor disagrees with its record (a SCOPED+ record's deep
review could be dropped by omitting the flavor). pre_eval.enabled: false writes a config-disabled FULL record that
--require-triage accepts while the stage stays disabled, instead of a dead end.

Fixed — Fable was reachable under cost-aware

The re-attempt escalation ladder and the reviewer lift ignored the stance. Both now stop at the stance's frontier
model (Opus under cost-aware); an explicitly pinned model is never escalated. A manifest without routing_stance
takes the project's stance (emitter, fast-path materializer and the resolver behind the residual dispatcher), and
without Codex the stance is kept instead of switching to claude-only, whose frontier is Fable. /v:init Step 4e
offers the advisor on Opus, or off, first under cost-aware.

Fixed — the plugin stayed silent on a plain request

The triage prompt hook spent its once-per-session chance before scoring, so a first message like "Read prd.md", a
greeting, or a scoring failure silenced it for the whole session. The session is now spent only by a SCOPED or FULL
decision; failures retry up to three times, a DIRECT decision does not silence the next request (capped at six), and
greetings and short read instructions are skipped. A project that is not set up gets one "run /v:init" line per
session, and a missing jq is reported once instead of silently disabling the hook. For a FULL tier the hook tells
Claude to start the pipeline unless the user says to skip it, rather than to "offer" it.

Added — install prerequisites and a lightweight budget profile

README and /v:init now name and probe what the flow needs: upstream Superpowers, jq, PyYAML for the interpreter
Engine C uses, and a first commit. /v:init offers a lightweight profile built from existing knobs (stance
cost-aware, recon off, retry: {max_attempts: 1, escalate_reviewer: false} — now read from
.claude/compound-v.json, not only the manifest — advisor on Opus or off, pre-flight 1A skipped on greenfield work).
Its savings are not measured. worktree.baseRef: head is offered default-yes.

Fixed — smaller

• Every command and skill resolves the plugin root with a guarded block; the old fallback pointed $CV at the
  user's repository, which has no scripts/.
• /v:collect and the integration gate use the same inputs as Engine C's gate (pinned baseline, digest-checked
  snapshot, toolchain_artifacts).
• register-lane --cwd may be omitted; a repository with no first commit gets a message saying so.
• resume-prepare lists leftover_violations: paths that blocked a superseded direct attempt and are still in the
  working tree. It never forgives them; it says so, so the next attempt is not blocked on the same file unawares.
• The Trigger 1 reminder named scripts/compound-v-emit-preflight.py, a path only this repository has; it now
  names the installed plu

Full detail in the CHANGELOG.