v1.1.1 - AI-import inbox and LaTeX regression lock

Release v1.1.1 - the delivery-path wave.

The two contracts that gate a finished AI/coding-agent import before it may reach the question bank, plus the guides that teach the complete path.

• AI-import inbox (dq_questionbank.import_inbox, schema ai-inbox-batch/v1): channels deliver a batch directory; the receiver computes the manifest SHA-256 itself (a digest declared by the delivering channel is never trusted). register_batch anchors batch_id + manifest into a confirmation_digest, or leaves the batch blocked with findings (missing/undeclared question files, malformed ids, unknown verdicts). verify_receipt recomputes the anchor immediately before transfer, so any post-registration edit - tamper or drift - is detected. Human review verdicts (passed / fixed / rejected) are validated against the declared file list.
• LaTeX regression lock (dq_questionbank.latex_regression + nine checked-in cases): every case pins detect (expect_issue_types: double superscript/subscript, malformed \frac), count (expect_fix_count via the shared stats contract), and repair (expect_equals / expect_contains / expect_not_contains / expect_no_faults_after) in one assertion. Locked edges: coordinate pairs and intervals never merge; three-column relation chains rejoin pair by pair; differentials go upright only inside integrals. Run anywhere: python -m dq_questionbank.latex_regression.
• Guide: OCR + AI Coding import (docs/ocr-ai-coding-import.md): the six-stage delivery path end to end - OCR pass to markdown -> first import under the work-file contract -> model self cross-check against page images -> independent checker thread (pdf_postflight) -> inbox registration -> human review and anchored transfer.
• Guide: LaTeX discipline (docs/latex-regression.md): one shared detector for the correction panel and the quality scan, the detect-count-repair trinity, and the change ritual for locked cases.

Fix: filesystem paths can no longer leak into introspected signatures and the public API manifest (default argument resolved in the function body; the manifest test now rejects drive letters and home directories).

Full changelog: https://github.com/wzsisshadiao-crypto/dq-questionbank-core/blob/main/CHANGELOG.md

Install: pip install dq-questionbank-core==1.1.1

v1.1.0 - coding-agent import wave

Release v1.1.0 - the coding-agent import wave.

Six pure modules extracted from the production import pipeline, each with deterministic tests and documentation:

• Word macro question-ID expansion (dq_questionbank.word_macro_id): the insert box accepts short references (6, DLLG_2025_6, KY SX SF SEU 2025 4) and expands them against a subject/school/year memory. Whitespace equals underscores, comma batches expand all-or-nothing, legacy ranges pass through, and ambiguous input returns None so the caller hands the raw text to the server instead of guessing.
• Coding-agent work-file contract (dq_questionbank.coding_agent_workfile, schema coding-agent-workfile/v1): validate_work_file reports schema, duplicate-number, status, empty-text, and missing-note violations plus any pipeline-owned field the agent file must not carry; transition_work_status enforces the transcription loop without mutating its input; writes are atomic and reads tolerate a UTF-8 BOM. Guide: docs/coding-agent-import.md.
• Read-only PDF postflight scanner (dq_questionbank.pdf_postflight): staged qNN.json candidates are scanned for numbering gaps, duplicates, missing identity fields, declared-vs-computed content hashes, and optional manifest.json digest agreement - findings are data, not exceptions.
• Paper metadata contract (dq_questionbank.pdf_metadata): a volume fails closed unless subject, controlled-vocabulary question type, and source are explicit and every question agrees with the first row; deployments may pass their own vocabulary.
• Import tag identity (dq_questionbank.pdf_identity): normalize_import_tag canonicalizes AH2026/ah_2026; job_matches_tag accepts a job id only when it equals the tag or extends it with exactly one known runtime namespace.
• LaTeX compatibility repairs (dq_questionbank.latex_compat): the $A$，$\quad B$ PDF degradation footprint is restored only when both sides are relation expressions; bare differentials become upright \mathrm{d} only inside integral math spans; both repairs count their fixes.

Full changelog: https://github.com/wzsisshadiao-crypto/dq-questionbank-core/blob/main/CHANGELOG.md

Install: pip install dq-questionbank-core==1.1.0

v1.0.0 - first stable release

DQ QuestionBank Core v1.0.0 — first stable release

v1.0.0 marks the completion of the open-source migration plan: the five
planned capability batches are all shipped, the four architecture issues
(#1 schema migration, #2 quality contract, #5 review workflow, #6 Word
publishing) are closed, and the remaining open issues are the five
community-input fixture requests that exist to collect legally shareable
corpora from outside contributors.

What the project is

Database-neutral, dependency-light infrastructure for structured
educational questions: a normative JSON document model, deterministic
quality checks, a review-first import pipeline, reference storage
adapters, and managed Word publishing — all synthetic-data only, no
production content.

The road here (14 pre-releases, 37 merged PRs, 35 closed issues)

• Foundation — canonical schema, validation, format registry, CLI.
• Visual workspace batch — structured formula editing, active-field
  tracking, scroll sync, soft-delete recycle bin, keyboard activation.
• Import & review batch — six synthetic intake cases (JSON/CSV/DOCX/
  PDF-style), digest-bound candidate sessions, configurable mapping
  import, OMML→LaTeX adapter.
• Daily question work batch — revision-bound quality findings,
  deterministic LaTeX repair rules, asset-repair and formula-image
  evidence contracts, SQLite + filesystem reference storage, safe
  backup-and-restore drill.
• Word publishing batch — managed content controls, revision-bound
  envelopes, loopback bridge, bundled VBA macro, publishing envelope
  documentation.

New since v0.15.0

• Schema 1.1 with a safer migration framework (#1) — analysis
  promoted to a first-class content field; built-in migrate() with
  refuse-to-guess path selection; load_schema/validate_with_schema
  handle both 1.0 and 1.1; new SUPPORTED_SCHEMA_VERSIONS and
  LATEST_SCHEMA_VERSION exports; full guide in
  docs/schema-migrations.md.

Stability commitments from 1.0.0

• The public Python API recorded in docs/public-api-manifest.json is
  stable; changes require an explicit manifest regeneration with
  compatibility notes (enforced in CI).
• Schema 1.0 documents remain readable and writable; future schema
  changes ship with deterministic migrations.
• 243 tests across Python 3.10–3.12 in CI; 280 audited public files;
  deterministic wiki generation from docs/.

Known scope and next steps

• The 2026-08-23 file-level audit against OPEN_SOURCE_BOUNDARY.md
  confirmed all planned batches migrated and the public tree free of
  private copies; the remaining eligible-after-review extractions
  (review-safety gates, splitters, PDF toolchain, scaling guides, and
  frontend modules) are tracked in the public roadmap (#83) as
  #84–#100 and land as narrow PRs with synthetic tests.
• The five open community-input issues (#28, #30, #40, #43, #48) wait
  on external fixture contributions and are not blockers.
• 1.1 schema exchange is opt-in; consumers must advertise support.

v0.15.0 - gap-closing batch

What is new

The gap-closing batch: every issue opened during the August migration audit now ships.

• OMML-to-LaTeX import adapter (#56, dq_questionbank.omml_import): reads native Word math (m:oMath/m:oMathPara) from a DOCX with the standard library only and maps fractions, scripts, roots, delimiters, named functions, and n-ary operators to deterministic LaTeX. Unknown constructs are preserved as text and reported for manual review - never guessed. Symmetric with the Word publishing export direction.
• Asset-evidence contract for image repair (#57, dq_questionbank.asset_repair): AssetRepairProposal binds the question block, asset reference, current digest, and replacement digest. Binding enforces the storage adapters' safe-relative-path rules; previewing re-verifies both digests; there is deliberately no accept-and-write step in the core.
• Formula-image transcription candidates (#59, dq_questionbank.formula_images): digest-bound review records for rasterized formulas. The formula_image flag is the deterministic signal; human transcription records its contributor; missing or drifted evidence fails closed. No OCR engine, external service, or network access.
• Safe backup-and-restore drill (#60, examples/backup_restore_drill.py): backup, verify, restore, and compare phases with SHA-256 manifests over a caller-chosen workspace; verification fails closed on any mismatch, and a tampered backup never overwrites anything. Guide at docs/backup-restore.md.

Full changelog: v0.14.0...v0.15.0

v0.14.0 - workspace trust batch

What is new

The workspace-trust batch: destructive-looking operations become reversible and the extension contract gets a runnable safe example.

• Soft-delete Recycle Bin (#58): recycling a question keeps it in the canonical payload - collection exports keep it until permanent delete - restoring returns it to its original position, and permanent delete removes it for real while clearing derived selections and marking the workspace unsaved. Covered by a Node vm logic suite driving the real app.js state machine.
• Safe plugin-discovery example (#38): examples/plugin_discovery_demo.py lists installed entry points without loading any plugin code; the --opt-in-load flag is the only step that executes third-party registrars. Documentation pins the deterministic clean-environment output.
• Blank-cell table fixture (#37): an original synthetic table with one intentionally empty cell, proven to keep its shape and the empty-but-present cell through rendering and canonical round-trips.

Full changelog: v0.13.0...v0.14.0

v0.13.0 - frontend quality batch

What is new

The frontend-quality batch: three community-visible rough edges closed.

• Keyboard-friendly Review Center (#16): every review row is focusable with a descriptive aria-label, Enter or Space activates mark-reviewed / undo, child-button activation stays native, and the focus outline is visible in both themes.
• Table-and-math rendering regression fixture (#27): the real app.js runs in a Node vm context with a scripted KaTeX double - proving table rows and header cells survive rendering, standalone and table-cell formulas both go through KaTeX, and a rendering failure keeps the source visible while removing nothing.
• Source-year filter regression test (#9): the served control, the exact matching predicate, the bundled case's 2025 subset, and the unknown-year empty state.

Full changelog: v0.12.0...v0.13.0

v0.12.0 - reviewable candidate sessions

What is new

The public reviewable import-candidate-session contract completes the review boundary between source extraction and persistence (#5).

• ImportCandidateSession / ImportCandidate wrap the canonical digest-bound session documents in stable public dataclasses - from_session verifies the digest, to_session reproduces the document exactly, and typed views never drift from the wire format.
• Parser identity is retained next to source evidence (parser.identity), so a review UI can always show which extraction produced a candidate.
• Deterministic candidate revisions: candidates start at revision: 1; a reviewed edit bumps the revision and rebinds the question digest. Plain accepts and rejects never touch it.
• Explicit states: decide() moves candidates pending -> accepted/rejected and fails closed on stale digests or already-reviewed candidates; export_accepted() returns a canonical QuestionSet and never persists.
• Stable fixtures for pending, reviewed (accepted with edit), rejected, and exported states; profile-extension guidance in docs/review-sessions.md.

Safety

• No candidate state implies persistence or AI approval.
• No private prompt, provider wiring, source document, or production schema is included.

Full changelog: v0.11.0...v0.12.0

v0.11.0 - writable reference SQLite storage

What is new

The first writable SQLite storage reference example, closing the storage gap flagged in #2.

• SqliteStorageAdapter implements the public StorageAdapter save/load contract using only the standard-library sqlite3 module - one canonical-JSON row per question-set identifier.
• Deterministic persistence: sorted canonical JSON, one transaction per save. Saving an existing identifier replaces the row entirely (documented last-write-wins); no partial writes are observable.
• Fail-closed loads: unknown ids raise KeyError; deserialization to a different id raises ValueError. Identifier rules match the filesystem adapter.
• Runnable demo: python examples/sqlite_storage_demo.py --db demo.sqlite3 builds a disposable database from the bundled synthetic fixture and validates the round trip through the public validation API.
• contains / stored_ids helpers for deterministic state inspection; docs in docs/sqlite-storage.md.

Safety

• The core stays database-neutral: this is a reference example, not the default backend, and it never opens or migrates a production database.
• No SQLite artifact is committed; .gitignore already excludes *.db / *.sqlite3.

Full changelog: v0.10.0...v0.11.0

v0.10.0 - revision-bound quality findings

What is new

The public revision-bound quality-finding contract (dq_questionbank.quality_findings) - the boundary a Quality Center integrates against.

• QualityFinding binds a question id, an exact target field path (stem.blocks[1], choices[0].content), a rule id, the ruleset version (quality/1), and SHA-256 input fingerprints over every field the rule read - incorporating the cross-field design raised in the issue #6 discussion.
• detect_quality_findings runs the deterministic LaTeX rules over math blocks in stems, solutions, and choices; repairable sources carry preview-only repair data, and nothing is ever rewritten automatically.
• finding_state returns current/stale: changing any declared dependency makes a finding stale, while unrelated edits never do.
• judge_finding records human accept/reject decisions as a separate operation that fails closed with StaleFindingError on stale content, binding the finding's own fingerprint.
• Stable serialized fixtures for current, stale, accepted, and rejected findings, plus contract documentation in docs/quality-findings.md.

Safety

• Detection, judgment, and persistence remain three separate operations.
• Unknown serialized fields and unsupported severities/decisions fail closed.
• No private correction rules, learned feedback, or production content is included.

Full changelog: v0.9.0...v0.10.0

v0.9.0 - deterministic LaTeX repair rule set

What is new

The first deterministic LaTeX repair rule set: several bounded repairs now work together behind one preview contract instead of living as isolated ideas.

• latex-bare-function-names — sin x + cos x becomes \sin x + \cos x, with word-boundary checks so \sin, assistant, and sinuous never match; \text{...}, \mathrm{...}, and \operatorname{...} regions are preserved.
• latex-delimiter-spacing — \left( x+1 \right) becomes \left(x+1\right); only the semantically empty whitespace directly inside the delimiter pair is dropped.
• latex-operator-spacing — x + 1 becomes x + 1 by collapsing two or more spaces next to + - = < >; single or missing spacing is never introduced or removed, and \text{a + b} stays verbatim.
• repair_latex_source composes the three new rules plus the existing latex-missing-closing-brace repair in a fixed order and reports every applied rule ID on applied_rules.

Safety

• Mismatched plain delimiters such as (x+1] are never rewritten; they return the latex-mismatched-delimiters manual-review finding with the source untouched.
• Every outcome keeps the original source visible next to the proposed latex, so a reviewer always sees the result before a question changes.
• New public API symbols are recorded in the stable API manifest; the public-tree audit now covers 212 files.

Full changelog: v0.8.0...v0.9.0