v1.2.1

This release covers everything since v1.0.0: the significance floor + collapsed digest (1.1.0), Word (.docx) and Markdown support (1.2.0), and the audited design/implementation gap fixes (1.2.1).

[1.2.1] - 2026-06-12

A promise-vs-implementation audit of every doc against the code (the same gap
class the 1.1.0 significance floor closed). Fixes, in order of weight:

Fixed

• ledger.js set now persists --tally (JSON-parsed). The protocol says
  "store tally and escalated on the row" and recall Mode B's consensus filter
  reads tally.valid, but the CLI's field whitelist silently dropped the flag --
  so the pre-edit consensus spot-check could silently select zero rows. Other
  unknown flags keep today's ignore semantics.
• meaning-audit seam corrected in review-engine-v3.md: the workflow
  returns {verdicts, arc} (the doc said anchor_verdicts), and a new SEAM 13
  spells out how the orchestrator enriches need_audit anchor ids into the full
  objects the workflow consumes (frozen_text from spine.json, support texts from
  anchor-diff). Following the old wording skipped the frozen-anchor revert gate.
• Auto compliance guard wired: submission-compliance.md promised that an
  edit pushing past the page limit or breaking anonymization is blocked in auto;
  the loop never invoked the checker. auto-mode.md and the EDIT-SAFETY chain
  now run compliance-check.js --pages <count> after compile-guard when a
  constraints file exists; a new blocker reverts + queues. No constraints file,
  no guard.
• Oscillation detection honestly retired: described across
  AUTO_MODE_DESIGN.md as a built deterministic guard, it was never
  implemented. Now annotated as superseded by the clerk re-raise merge + the
  rounds-touched cap + applied-quiescence; the phantom reference in
  decompose.js's header is gone.
• Doc/protocol drift swept: config-template no longer calls auto "DESIGNED, not
  built" (it shipped) and documents mode: auto + intensity; the v3 persona
  default (runtime assign-reviewers, lenses as fallback) replaces the legacy
  default; writing-toolkit.md keys on valid-fixable (v3) instead of the v2
  agreed-to-fix; spine.md documents the bare-array stdin shape spine.js
  actually reads; ledger-schema.md re-documents rounds_touched (derived from
  the journal, row field reserved) and drafted_patch (optional, module API)
  honestly; the clerk header no longer claims re-trial rows are carried;
  machine-specific environment facts removed from submission-compliance.md;
  "reference style" moved off the deterministic-checks list it never shipped on;
  AUTO_MODE_DESIGN.md's status header records the 2026-06-05 end-to-end run
  and strikes two long-done build tasks; the legacy review-panel workflow's
  reviewer/verifier prompts gain the ISOLATION line hard rule 2 promises.
• Public copy caught up: READMEs document the significance floor + collapsed
  digest view, the full ledger.js CLI (incl. mode/floor), the 1.2.0
  scripts (extract-docx, rekey), and the npm run doctor health check
  (previously documented nowhere); overview.html stops listing the shipped
  plugin marketplace as a TODO and de-LaTeXes the direct-edit card.

[1.2.0] - 2026-06-12

Word (.docx) and Markdown support, from real user feedback. The LaTeX path is
byte-identical to 1.1.0: default flags and absent options behave exactly as before.

Added

• Word (.docx) support via one-time extraction. New dependency-free
  scripts/extract-docx.js converts a .docx ONCE into a Markdown working copy
  at .paper-review/<basename>.md plus a SEPARATE machine-readable
  extraction-report.json (per-feature counters and a char-mass audit, so
  anything dropped or degraded surfaces as a number, never as silent loss; the
  report is never embedded in the working copy). The original Word file is never
  modified; from extraction on, the working copy IS the manuscript for every
  rule and gate. Fail-loud matrix with named reason codes (legacy .doc OLE2,
  encrypted / zip64 / unsupported-method zips, UTF-16 document.xml, missing
  main part, not-a-zip). Refuses to overwrite an existing working copy without
  --force; a changed original (sha256 mismatch vs ledger meta) stops and asks
  instead of silently re-extracting.
• Markdown as a first-class multi-round format. decompose.js markdown mode
  (counter-based section paths from ATX headings, fence-aware, no %-comment
  stripping, so "12% over baseline" stays intact); compile-guard.js routes
  non-.tex working copies in-script to compiled:null plus a markdown lint
  (an honest UNKNOWN, never latexmk on a .md); compliance-check.js runs only
  its format-neutral subset on Markdown with an explicit skipped_checks list
  (no false section-missing majors).
• Round-end rekey + passage-alias map (both formats). New scripts/rekey.js
  re-links open ledger rows whose passage_id no longer resolves after edits
  (via the journaled after-text, else the row's evidence_anchor) and maintains
  .paper-review/passage-aliases.json, consulted by the journal.js cap
  functions. This restores the clerk merge key and the rounds-touched cap after
  edits move first-words anchors; unresolved rekeys are recall-safe and listed.
• apply-patch.js --guard-paragraphs (opt-in). Rejects a patch whose
  before/after blank-line paragraph counts differ (a paragraph split/merge
  cascades passage ordinals). Mandated by protocol on Markdown working copies;
  default OFF on .tex (LaTeX behavior unchanged).
• Intake format gate. SKILL.md, docs/AGENT-GUIDE.md,
  references/review-engine-v3.md (new step 0), and the auto-mode pre-loop now
  route every manuscript by format: .tex native, .md/.markdown/.txt
  native text path, .docx one-time extraction, anything else explicitly
  unsupported with export suggestions. No more silent degradation.

Fixed

• Drift-bound doc correction. docs/AUTO_MODE_DESIGN.md and the engine docs
  no longer imply a passage's anchor survives edits: a first-stable-words anchor
  mutates when a paragraph's opening words are edited (and Markdown has no
  \label-immune subset). The documented mechanism is now the real one:
  rounds-touched cap + round-end rekey + alias map, failure direction
  recall-safe.
• configs/config-template.md ledger default corrected from LEDGER.md to
  LEDGER.json (the JSON is the machine source of truth; the .md is a
  rendered view).

Notes

• Rejected on record (rationale in the docs/AUTO_MODE_DESIGN.md changelog):
  extraction-time injected HTML-comment anchors, a hash-window change to the
  anchor scheme, and a pandoc extraction fork (pandoc cannot emit the honesty
  report and yields machine-dependent working copies). Converting your own
  document to .md and handing that in directly remains fully supported.

[1.1.0] - 2026-06-12

The trivia-flood fix (F3), from real user feedback: "阻断 AI 去关注非常细微没有价值的
问题" (stop the AI from chasing tiny, worthless issues). Backward-compatible; no
schema bump, existing ledgers render identically until a mode is set.

Added

• The significance floor, now in code. node scripts/ledger.js floor <ledger.json> returns {fixable, excluded}: the drafter's input is exactly the
  valid-fixable MAJOR rows, and any valid-fixable non-major is excluded with its id
  reported (read-only, never silent). This is the floor references/auto-mode.md
  had promised; it is now the normative builder of the drafter's fixable set
  (review-engine-v3.md step 13 / SEAM 4). A polish item escalated to trial is
  promoted to significance: major as part of the escalation contract, so a
  later valid-fixable verdict on it passes the floor.
• Collapsed ledger view. LEDGER.json meta gains an optional display_mode
  (show|collapse; absent = show, the previous behavior byte-for-byte). In
  collapse, LEDGER.md keeps majors as itemized table rows and folds minors
  into a "Minor digest": open/queued minors one compact line each (a pending
  decision is never hidden), terminal minors as per-status count lines, plus a
  never-drop footer. Render-only -- counts, the completion gate, statuses, and
  routing are untouched, and full detail stays in LEDGER.json. New commands:
  node scripts/ledger.js mode <ledger.json> <show|collapse> and
  node scripts/ledger.js init ... --display <show|collapse>. Auto mode
  initializes collapsed; review mode keeps the flat table by default.

Notes

• The minor/mechanical fixes themselves still happen (the polish track is
  unchanged); the flood is treated at the presentation layer, never by dropping
  work. Issues are never silently dropped.
• reason_code: batched-nit remains in the schema as RESERVED: the composite-
  packing design it anticipated was evaluated against real run data and rejected
  (see docs/AUTO_MODE_DESIGN.md changelog, 2026-06-12, which also records the
  formal override of the 2026-06-10 design-debate archive's ship recommendation).