emrg: journal_prompt — citation authenticity verification + >=100 references gate

[argszero]

Summary

Implements host rant 2026-09-10T11:28:43: the journal task prompt (emrg/server/journal_prompt.md) gains two citation requirements — citation authenticity verification (anti-fabrication) and a ≥100-reference threshold (with in-text coverage).

Rationale from the rant: the prompt previously had only scattered anti-fabrication wording ("Do not fabricate data or citations") and a minimum-3-comparison requirement — no systematic citation verification flow and no volume bar.

Changes

Author side — Submission quality bar (now 13 items):

• emrg: fix remaining bare excepts on wait_closed() and add registry tests #12 Citation authenticity verification (mandatory): every reference must be verified against a real external record before submitting. DOI present → Crossref works/<doi> and compare title/authors/year; no DOI → title search via Crossref query.bibliographic or the arXiv API. Results are written to papers/issue-<N>/reference-check.md (per-entry: key → method → result → real record found) and submitted with the manuscript. Unverifiable entries must be deleted or replaced — never submitted.
• emrg: add bash_tool and base tool tests #13 ≥100 references, all genuinely cited: a hard, level-independent threshold (case studies included). Bib entries never cited in the body are padding and do not count. Stacks with emrg: fix dead 256-color branch in style_to_sgr, add test #3 to form the citation gradient: 3 comparisons (floor) → 100 references (volume) → authenticity verification (quality).
• Manuscript layout now lists reference-check.md as a committed artifact.

Editor side — Review quality bar (now 13 items):

• emrg: fix remaining bare excepts on wait_closed() and add registry tests #12 Independent citation spot-check: the editor must not trust the author's self-report — sample several references (≥1 DOI-less/suspicious entry) and re-verify via Crossref/arXiv. A fabricated or unverifiable citation is a major concern and can alone justify REJECT.
• emrg: add bash_tool and base tool tests #13 Reference-count + in-text coverage check: total ≥100 and per-entry body coverage; below threshold or many uncited entries → return for revision.

Wiring / consistency:

• Citation verification row added to both review comment templates (editor + author Phase Review-Other), mirroring the Significance check duplication: sampled <n> / fabricated <m> / unverifiable <k> — total references <T> (≥100 required), uncited entries <u>.
• Editor Phase A triage (completeness step) now verifies the citation gate (≥100 + coverage + reference-check.md present) before moving to in-review; the independent spot-check still happens at review time.
• Common Rules emrg: consolidate duplicate SessionMemoryStore imports in daemon.py #8 extended so the bar is visible outside the phase sections.
• Verification method is kept declarative (API endpoints + judgement criteria only, no embedded scripts), per the existing session guidance on keeping prompt files declarative.

Tests

5 new content guards in tests/test_journal_prompt.py (18 total in that file):

• test_editor_review_bar_requires_citation_authenticity_check
• test_reference_count_threshold_and_in_text_coverage
• test_author_side_citation_integrity_requirements
• test_citation_gate_wired_into_triage_and_common_rules
• test_citation_verification_row_in_both_review_templates (asserts template parity == 2)

Verification

• uv run pytest tests/ -q → 1241 passed, 1 skipped (Agent.md count synced 1237 → 1242 collected)
• uv run python -c "from emrg.client.app import run_client" → import OK
• uv run python -m emrg --help → CLI OK
• Jinja2 render check for both roles (editor / author) → renders cleanly, both carry the Citation verification row
• Both quality-bar lists verified for continuous numbering (1–13) with no duplicates

[argszero]

✅ LGTM — cycle 20260910-114658

Reviewed 0acd344 against a729e3a (3 files, +68/−3) and independently verified the branch content (not the working tree).

Rant coverage (2026-09-10T11:28:43) — both requirements land at every requested anchor:

• Author-side Submission quality bar #12: mandatory verification of every reference before submitting (DOI → Crossref works/<doi> title/authors/year compare; no DOI → query.bibliographic title search or arXiv), results written to papers/issue-<N>/reference-check.md and shipped with the manuscript; unverifiable entries deleted/replaced.
• Author-side #13: ≥100 references, each genuinely cited in the body (padding entries excluded), hard level-independent threshold incl. case studies; explicitly stacked with #3 into the citation gradient.
• Editor-side Review quality bar #12/#13: independent spot-check (never trusting the author's self-report, ≥1 DOI-less sample) with a fabricated/unverifiable citation as a stand-alone REJECT basis; plus count + in-text coverage check.
• Review comment template row added to both templates (parity verified == 2), the triage completeness step carries the citation gate, and Common Rules #8 is synced so the bar is visible outside the phase sections.

Verification performed:

• Both quality-bar lists re-extracted from the committed blob → numbered 1–13, continuous, no duplicates; the existing 11 items' semantics are untouched (pure appends).
• Verification method kept declarative (API endpoints + judgement criteria only, no embedded scripts) — matches the rant's explicit instruction.
• 5 new content guards, including count-based template-parity assertions; uv run pytest tests/ -q green locally (1241 passed / 1 skipped); Agent.md count synced 1237 → 1242.
• Jinja2 renders cleanly for both editor and author roles with the new row present.
• CI: test pass (1m59s) + test-windows pass (3m2s) on head 0acd344.

No blocking issues found. LGTM 1/3 (author cycle 20260910-113202 excluded).

[argszero]

✅ LGTM — cycle 20260910-115454

Re-review of 0acd344 (head unchanged since LGTM 1/3) — content independently re-verified against the current branch head, not the working tree.

Re-verification (fresh this cycle, from origin/feature/journal-citation-integrity):

• Editor Review quality bar re-extracted → numbered [1..13] continuous: True
• Author Submission quality bar re-extracted → numbered [1..13] continuous: True
• Citation verification row count == 2 (editor + author review templates — parity holds)
• Crossref DOI endpoint (api.crossref.org/works/<doi>) present; declarative style preserved (no embedded scripts)
• reference-check.md referenced 4× — the authenticity report is wired as a committed submission artifact
• Diff vs a729e3a: 3 files, +68/−3 (prompt + tests + Agent.md count sync)

CI (head 0acd344): test pass (1m59s) + test-windows pass (3m2s); PR MERGEABLE / CLEAN.

No blocking issues. Still satisfies rant 2026-09-10T11:28:43 (both requirements at every requested anchor: author bar #12/#13, editor bar #12/#13, both review templates, triage gate, Common Rules #8). LGTM 2/3.

[argszero]

✅ LGTM — cycle 20260910-120201

Third independent review of 0acd344 (head unchanged since LGTM 1/3) — content re-verified from the committed branch blob, not the working tree.

Independent verification (fresh this cycle, isolated detached worktree at 0acd344):

Check	Result
Editor Review quality bar	numbered [1..13] continuous: True
Author Submission quality bar	numbered [1..13] continuous: True
Citation verification review-template rows (raw)	2 — parity across editor + author templates holds
Per-role Jinja2 render (editor / author)	both OK (26090 / 32256 chars), 1 citation row each — no render errors
Crossref refs	DOI endpoint api.crossref.org/works/<doi> + query.bibliographic title search + arXiv fallback all present
reference-check.md wiring	referenced 4×, declared as a shipped submission artifact
PR-branch test suite	tests/test_journal_prompt.py → 18 passed (5 new guards, incl. count-based template-parity assertion)
Branch collection	1242 tests collected (matches the synced Agent.md count)
CI on head	test pass (1m59s) + test-windows pass (3m2s) — PR MERGEABLE / CLEAN

Rant coverage (2026-09-10T11:28:43) — both requirements land at every requested anchor: author bar #12 (mandatory per-reference verification → reference-check.md, delete/replace unverifiable entries) and #13 (≥100 references, each genuinely cited, hard level-independent threshold); editor bar #12 (independent spot-check, never trusting the author's self-report, ≥1 DOI-less sample; fabrication alone justifies REJECT) and #13 (count + in-text coverage); citation gate wired into triage; Common Rules #8 synced.

Style check: the verification method stays declarative (API endpoints + judgement criteria), with no embedded ad-hoc scripts — exactly as the rant requires. Existing bar items are untouched (pure appends).

Housekeeping: local uv.lock churn from the test run was reverted; all verification artifacts (worktree + temp files) removed — working tree clean.

No blocking issues found. LGTM 3/3 (author cycle 20260910-113202, plus voting cycles 20260910-114658 and 20260910-115454, are all distinct).