v4.27.0 — 118 rounds of using the product instead of reading it

Fixed — 118 rounds of using the product instead of reading it

The published version throughout this work was 4.23.1. Anyone who installed
from PyPI in that window got a build predating 4.24.0, where a run carrying a
.csv, .json or .docx source could not be published at all: reference
curation stripped the local-artifact label and the reference gate then
hard-blocked on a citation it no longer recognised. That fix, and the three
releases of work since, reach an installing reader only now. Cutting this
release is the point of it.

The 118 commits below were all found the same way — by running the product on a
real document until it stopped, then fixing what stopped it. Halfway through,
the framing changed from "I am already inside this workflow, where does it
leak" to "I have just found this repository, where does it stop me", and the
last two dozen entries come from that second question.

Fixed — the files people actually have

• Big5/cp950 sources open now. Every source ever tested was UTF-8 because
  the tester wrote them; Excel on a Traditional Chinese Windows saves CSV as
  Big5 by default. Three separate reading paths were hard-coded to UTF-8 and
  now share one decoder (utf-8-sig → utf-8 → cp950 → gb18030). The
  worst of the three never failed: parse_code read with errors="replace",
  so a Big5 file returned success: True with a body of U+FFFD, and the
  mojibake was admissible evidence. No latin-1 catch-all was added — it
  decodes anything, which turns a broken file into silent nonsense.
• A workbook is more than its first sheet, and its title cell is not a
  column name. openpyxl is also a declared dependency now; it never was, and
  a clean install accepted .xlsx as supported and then could not open it.
• Word and PDF tables survive the trip. Merged header cells dropped one of
  the two readings beneath them, cells past the last column vanished without a
  word, a table continued onto page two named its columns after a reading, and
  an equation was dropped on the way in and rewritten on the way back. Table
  rows extracted from .docx/.pdf are citable one row at a time, as CSV rows
  always were.
• A table pasted out of a spreadsheet is a table. Tab-separated text glued
  into a note was swallowed by the paragraph above it and registered as one
  qualitative block — six readings, no chart, no statistical claim.
• Broken attachments are described in your vocabulary, not this build's. An
  empty .md was reported as "agent fallback parser is not implemented in the
  local MVP", a scanned handout was refused as if the file were corrupt, and
  three unreadable attachments were reported one run at a time — three
  round-trips for information the first parse already had.

Fixed — Chinese was being checked in English

Thirteen rounds of Chinese documents ran through checks whose vocabulary was
Latin-only, so they passed by not applying.

• A number spelled in Chinese escaped the numeric check; a fabricated
  quotation was only scanned in English; the banned-phrase list had no Chinese
  entries; three figure checks and the duplicate-caption check never fired on a
  Chinese report; a Chinese claim citing English evidence got no vocabulary
  check at all.
• Reference curation skipped Chinese reports entirely — GB/T style returned
  before the filter ran, so every Chinese bibliography shipped uncurated,
  including the report's own data.csv.
• Revising a Chinese report renamed 摘要 to Abstract, and a Chinese heading was
  short enough to be discarded as noise; 備審資料 could not select the profile
  built for 備審資料; a Chinese header matched no term at all, so a trial
  counter was chosen as a chart axis; a Chinese paper could never be classified
  as literature, while the warning kept asking the author to attach the
  literature they had attached.
• CJK typography normalization was rewriting figure file paths — correct for
  prose, destructive for the one thing in the document that is not prose.

Fixed — statistics with no referent

The derived-statistics layer computes what a grader looks for; each new
inference had to be asked what it would conclude on a different shape of data.

• A column held constant is a controlled variable, not an axis; three steady
  columns were multiplied into a statistic by accident; a curve was fitted
  against a column that never changed, and a measurement was regressed on its
  own row number. A rated/nominal column is now recognised as the reference
  curve it is, in both languages.
• Evidence that certifies itself is no longer accepted: a section heading
  could ground the claim that restated it, and a transcript question could
  ground the claim it was asking about. The same file attached twice became
  evidence twice, and padding a thin source set with duplicates cleared the
  source-base bar.
• A table nobody filled in was read as measurements; two of three thermocouples
  were destroyed at ingestion; a file name was cited as if it named the authors;
  a vault's bookkeeping and a note the author had hidden from themselves were
  both printed into the report.

Fixed — revising your own report

• Revising deleted the figures and left their captions; took the author's name
  off the cover; and put an underscore through every numbered heading.
• A report being revised could not cite the measurements printed in it, and
  revising it destroyed its tables.
• Removing a section that held a figure could only pass by lying about it; a
  revision aimed at the wrong section id lost the sentence in silence;
  renumbering a heading changed which section it was; and adding a missed trial
  discarded every citation below it.

Fixed — the package you actually send

• The delivery bundle shipped the scaffold beside the document, and the
  delivery summary answered "which file do I send" with a working copy. The
  client-readable note was neither readable nor the answer, and was written in
  the wrong language.
• A clean report came back marked "review" over two optional tools nobody had
  asked for.
• A source that moved before publish was dropped from the bundle in silence, a
  bundle could ship a source that no longer says what the report quotes, and of
  two same-named sources only one was packaged.
• Thirteen [1] markers shipped over an empty bibliography; a raw [FIGURE:]
  placeholder reached the final document; a figure could disappear from it
  without a word.
• A delivered report no longer carries the machine that made it. pandoc
  writes the image's source path into the picture's description, and a run
  directory is named after the prompt, so every report with a chart shipped an
  absolute local path and the author's prompt inside the .docx, where Word
  shows them in the picture's alt-text pane. The caption was already computed
  and already correct on the neighbouring element; POST_RENDER_REPAIR now
  hands it to the one that was holding the path. Found by scanning the sample
  document below before committing it, not by reading the code.
• Your own template works on a finished report. --reference-docx was
  refused on any report that had already rendered, because the guard demanded
  status validated and a delivered report is completed — strictly more.
  A course template's cover page also vanished without a word, and the
  template-fidelity report was checking the built-in template rather than the
  author's.

Fixed — the first hour with the tool

• prepare --output <dir> put the run where it was asked and then status
  could not find it; status itself reported one of the four things the run
  needed; diff could not compare two checkpoints, which is what it is for;
  remap-evidence rewrote three artifacts and reported touching none;
  invalidate-cache told the author to pass the flag they had just passed;
  diagnose told a brand-new job its revision had succeeded; and pointing at a
  folder was reported as a permissions problem.
• Task briefs contradicted the gates they describe: an abstract contract that
  was not this run's, a key nothing reads, and a hard rule labelled for a
  profile the run is not.
• An outline with seven empty sections cost seven publish attempts, one per
  round-trip; a blocked stage now says how to get unblocked; pandoc's list of
  dropped elements went to a truncated log line; and two of three
  evidence-policy checks only ever reached stderr, so a run that raised three
  warnings recorded one.
• Tables went out headed by raw column names, a stacked header made a sheet
  unreadable, and a chart that could not draw its own labels said nothing.

Fixed — claims this repository made about itself

• CI had been red for eighty-one commits while every round reported green.
  Python 3.11 could not byte-compile reference_verify (a backslash inside an
  f-string expression), openpyxl was undeclared, and the mcp guard probed a
  module the server does not import. Reporting green from one interpreter on one
  operating system is not reporting green.
• The dependency list the README installs had drifted from the real one; the
  first command a stranger runs after installing did not exist; and the
  credibility page quoted an adversarial run that is not the archived one. The
  archived numbers are now pinned by contract tests, as the test-count badge
  already was — including the count on docs/EVIDENCE.md, which had rotted to
  496 against a suite of 783.
• examples/source_to_report.py was added: three files and one sentence in, a
  finished DOCX out — the thing the tool is for, which the examples directory
  did not previously demonstrate. The README was cut from 534 lines to 134.

Added — what a stranger sees before running anything

• examples/output/ holds the document examples/source_to_report.py
  produces, committed: the .docx itself plus the client-readable QA note that
  says, claim by claim, which source row each sentence rests on. A contract test
  keeps it a real deliverable — headings, a Word table, a chart, no local path —
  so the thing most readers judge the tool by cannot quietly rot.
• The README leads with the agent path. Installing the skill and asking for
  the report in your own words comes first; the CLI follows as the scripted
  route. The MCP section now says plainly that it exposes the gates, not the
  whole pipeline, which is what a reader discovers three commands later anyway.
• The Colab notebook runs the whole path, not just verify(): install,
  three source files, a rendered .docx you can download, then the gate on its
  own. It installs from the cloned repository rather than PyPI so the notebook
  always matches the code beside it.