feat(svm): write the svg through a writer that escapes, and log what we drop

[andiwand]

🤖 Generated with Claude Code

Stage 1 of #772 — the infrastructure the rest of the work stands on. The
staged plan is written down in src/odr/internal/svm/PLAN.md; this PR is the
base of the stack.

The problem

svm_to_svg.cpp wrote its markup straight into the stream. Two consequences:

• Text was not escaped. write_text did out << text. An &, < or >
  in a chart label makes the svg malformed, and a malformed svg renders as
  nothing — the browser drops the whole image, not the one label. (Improve SVM to SVG conversion #772
  defect 1.)
• Nothing was logged. Every unhandled action carried a // TODO log, and
  write_image_src caught every exception and fell back to emitting the raw
  .svm bytes as a data url that no browser renders. A file we drew correctly
  and a file we drew as a blank rectangle looked exactly alike. (Improve SVM to SVG conversion #772 defect
  10.)

What is here

• svg::SvgWriter (internal/svg/svg_writer.*), the counterpart to
  html::HtmlWriter: elements, attributes, style declarations and text, all
  escaped, with the style attribute accumulated per element so attributes and
  style can be written in any order. A ; in a style value is dropped rather
  than allowed to open a declaration of its own.
• One escaper and one number formatter, shared with html. The escape lives
  in util::xml::escape_text / escape_attribute — one pass, dropping the
  control characters xml 1.0 cannot carry — and html::escape_attribute, four
  replace_all passes over a by-value string, is gone in favour of it.
  html::escape_text stays in html: its   and   are
  presentation, not escaping, and   is undefined in xml. Numbers go
  through util::number::to_string_significant, which already documented
  "never in scientific notation, which CSS and SVG lengths do not accept"; the
  classic locale it needed (a german one writes 1,5 into a coordinate) moved
  in there, so Measure::to_string gets it too.
• A logger through the translator. Translator::svg becomes
  svm::translate_to_svg(file, out, logger) and reports every action it skips
  by name (action_type_name in svm_format). Getting a logger there needed
  one: abstract::HtmlService gained logger(), so a view reaches the sink
  its service was built with, and WritingState carries it down to
  translate_image_src.
• write_image_src asks the file type before trying an svm translation,
  instead of attempting one on every image and catching. That is what makes a
  failure worth a warning rather than noise on every png.
• write_style(out, context, int) with its magic 0/1/2 is an enum now
  (Improve SVM to SVG conversion #772 defect 9).
• Docs: svm/README.md keeps the feature checklist, now checked against
  the code and completed — the shapes, fills, clipping and state handling that
  were never listed, and the references it was missing. svm/AGENTS.md (the
  format, the conventions) and svm/PLAN.md (the stages) are new; a row in the
  root AGENTS.md; a section in svg/AGENTS.md for the writer.

Tests

svm_test.cpp builds its input as bytes, inline (SvmBuilder), so an
action gets a test without a fixture — the harness the following stages need.
Six new cases: the empty file, a rectangle, text, text escaping, attribute
escaping, and that an unimplemented action is skipped by its own length so the
one after it still reads.

Output is unchanged

Rendered the four svm fixtures and diffed the decoded svg against main's:

fixture	before	after
odr-public/svm/chart-1.svm	6658 B	identical
odr-public/svm/table-1.svm	9358 B	identical
odr-private/svm/test.svm	18643 B	identical
odr-private/svm/Vyplaty.svm	21913 B	identical

The formatter can only part from operator<< where a value would print with
an exponent — |v| ≥ 1e6, or 0 < |v| < 1e-4 — and no coordinate in the four
fixtures is in that range. Where it does bite it is the fix: 1051938 rather
than 1.05194e+06, which is not a css length at all.

The full suite passes (1285 tests), HtmlOutputTests included, so the shared
escaper changes no reference rendering either. Decoding all 1144 generated
svgs in the reference output shows none of them currently malformed, so the
escaping fix changes no rendering today — the bug is latent in the corpus, not
absent from the wild.

Known gap, recorded rather than fixed

Escaping is only half of #772 defect 1. read_string_with_encoding hands back
the file's own bytes for every encoding but UCS2, so a latin-1 label emits
invalid utf-8 and an xml parser refuses the document just as hard as an
unescaped & made it. README.md records it and stage 3 of PLAN.md picks
it up.