When a specialist reviewer has ≥1 prior review record for THIS project,
paper_reviewer.py now prepends the shared re-review block (from
agents/prompts/_shared/rereview_block.md) to the user prompt, with the
specialist's most-recent prior action_items substituted in. The block
instructs the LLM to apply the two-question protocol (FR-014/015/016)
instead of generating a fresh critique.

A specialist with NO prior records continues to use the full-critique
prompt (FR-017). This is the per-specialist toggle from clarification
session Q2.

Changes:
- src/llmxive/state/reviews.py: prior_reviews_for_specialist() filters
  list_for() output to one specialist + sorts by reviewed_at ascending.
- src/llmxive/agents/paper_reviewer.py build_messages: when prior reviews
  exist FOR THIS specialist, render the shared snippet with the most-
  recent prior's action_items as YAML, prepend it to the user prompt.
- contracts/review-record.schema.yaml: action_items array added so old-
  record-validation doesn't reject the new field on serialization.
- tests/unit/test_rereview_per_specialist_toggle.py: 7 new tests covering
  per-specialist filtering, sort order, snippet presence, no-priors path.

Full unit suite (458 tests, +7) still passes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds the operator escape hatch and scheduler skip rules required by
spec 012:

- src/llmxive/cli.py: new subcommand `llmxive project unblock <PROJ-ID>`
  (FR-023). Refuses to no-op-unblock: requires the most-recent
  state/revisions/<PROJ-ID>/round-N.yaml file to be modified AFTER the
  project's recorded updated_at (mtime check). Transitions to
  PAPER_REVIEW by default; --to-minor transitions to PAPER_MINOR_REVISION.
- src/llmxive/pipeline/scheduler.py: PAPER_REVISION_IN_PROGRESS,
  READY_FOR_IMPLEMENTATION, and PAPER_REVISION_BLOCKED added to
  _NEVER_PICK. FR-009's idempotency rule: while a project is being
  planned, the regular scheduler MUST NOT re-trigger work on it. The
  ready/blocked states are owned by dedicated agents (implementer +
  human respectively), not the regular tick-scheduler.
- tests/unit/test_cli_project_unblock.py: 5 tests covering happy path,
  --to-minor flag, no-op-unblock refusal, wrong-stage refusal, missing
  round-file refusal.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

3 integration tests in tests/integration/test_revision_in_progress_idempotency.py:
- verify the three spec-012 stages are in _NEVER_PICK
- verify a runnable project is preferred over an in-progress one
- verify the scheduler returns None when every project is in a NEVER_PICK state

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

When a home-grown paper enters PAPER_REVIEW with writing/science action
items (no fatal), advancement.py now transitions the project to
PAPER_REVISION_IN_PROGRESS and invokes revision_planner.run_revision_pipeline.
The planner produces a full revision-spec directory under
specs/auto-revisions/<PROJ-ID>/round-<N>/ containing spec.md, plan.md,
tasks.md, analyze-report.md, and result.yaml.

Implementation is DETERMINISTIC (v1): each of the 5 stage outputs is
generated directly from the consolidated action items (no LLM call).
The spec/plan/tasks artifacts are concrete enough that an implementer
agent can pick up the revision_spec_path and execute. A follow-up PR
replaces the deterministic generation with the full LLM-driven speckit
pipeline (speckit-{specify,clarify,plan,tasks,analyze}).

Public API contract is stable across v1 (deterministic) and v2 (LLM-driven):
  run_revision_pipeline(project_id, action_items, *, revision_kind, repo_root)
    -> RevisionSpecResult{revision_spec_path, final_outcome, stage_results, ...}

Defensive checks:
  - ArxivIntakeError on arxiv-intake projects (advancement.py routes
    them through upstream_feedback instead).
  - RevisionPlanningError on FS/schema failures.
  - On either error, advancement.py transitions to PAPER_REVISION_BLOCKED
    so the operator notices.

state/revisions/index.yaml is also updated atomically so an implementer
agent can discover ready-for-implementation projects without scanning
the filesystem.

8 new unit tests in tests/unit/test_revision_planner.py cover the
5-artifact generation, action-item-to-task mapping, arxiv-intake
guardrail, science vs writing kinds, round-number incrementing, and
the index.yaml update.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… (T052)

- web_data.py: PAPER_REVISION_IN_PROGRESS, READY_FOR_IMPLEMENTATION,
  PAPER_REVISION_BLOCKED added to _PHASE_GROUP_BY_STAGE (all → paper_review
  phase). Without this, projects landing in the new states would be
  rendered as "blocked" (the fallback group), which is misleading.
- _project_to_entry payload gains revision_spec_path (links to the
  auto-planned revision spec dir when stage == READY_FOR_IMPLEMENTATION)
  and upstream_feedback (summary of the arxiv-intake annotation).
- _upstream_feedback_summary() reads upstream_feedback.yaml and returns
  {schema_version, round_count, latest_verdict_class, latest_action_item_count}.
  None when the file is absent (most projects).

Regenerates web/data/projects.json with the new fields.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…gate (T050)

Adds the end-to-end convergence test required by SC-001 / T050. The test
covers the three terminal outcomes:
  - All specialists accept → PAPER_ACCEPTED
  - Writing-class action items → PAPER_REVISION_IN_PROGRESS → 5 artifacts
    + READY_FOR_IMPLEMENTATION
  - Fatal-class action items → BRAINSTORMED + rejection rationale appended
    to the idea record

Gated on LLMXIVE_REAL_TESTS=1 per the real-call test convention. The test
exercises pure-Python logic + real filesystem state (no Dartmouth calls
needed; the deterministic revision_planner emits artifacts directly).

ALSO fixes a defensive bug in _all_specialists_accept_most_recent:
previously, when `required` was empty (registry-load failure), the gate
trivially returned True — which meant any non-accept review on an
unconfigured registry would be incorrectly routed to PAPER_ACCEPTED.

New behavior:
  - empty required + no records → False  (unconfigured; refuse to advance)
  - empty required + all-accept records → True  (every reviewer that
    recorded a verdict accepted; vacuously OK)
  - empty required + any non-accept → False  (severity branch takes over)
  - non-empty required + records → standard per-specialist most-recent check

Two unit tests added in test_advancement_convergence.py to lock the new
behavior in (replacing the prior single test_empty_required_gate_passes_trivially).

Full unit suite (463+e2e) passes locally.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Updates the "How it works → The paper pipeline" section to describe the
spec-012 convergence pipeline (structured action items, most-recent
verdict gate, three-way severity routing, per-specialist re-review
protocol, and arxiv-intake guardrail).

Closes the last remaining task in the spec-012 task list (T053). With
this commit, all 55 of 55 tasks are now landed on PR #198.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>