Collapse the duplicated enum tables and XML helpers the persistence layer maintains twice (ponytail audit) #563

A repo-wide over-engineering audit (244k tracked source lines, all 13
FetchContent deps, 9 CMake options, 14 CI jobs) found the tree broadly
disciplined: no dead public API (all 262 RM_API functions have a
non-test caller), no unused dependency, no dead build flag, one
orphaned tracked image.

What it did find is a cluster of deliberate duplication, most of it in the
persistence layer: the OpenDRIVE and OpenSCENARIO reader/writer pairs each keep
their own copy of the same XML helpers and their own one-directional
enum↔string table — several with comments explicitly defending the copy
against a divergence that has not happened in the project's lifetime.

That duplication is not cosmetic. #476 was exactly this class of defect —
the writer rewrote lane/mark enums into different semantics than the reader
parsed, because the two tables were maintained separately. The probe on that
issue found 9 corruptions where the issue body listed 3. Collapsing each pair
to one table makes the class unrepresentable rather than merely fixed.

Scope — two sequential PRs, no behavior change

PR-A — kernel dedup (≈ -235 lines)

Three invariants A1 must not violate, each preserving today's behaviour:

1. Write-fallbacks are not table rows. LaneType::Other → "none",
   ObjectType::Other → "none", RoadMarkType::Other → "solid" are the
   writer's fallback argument. As table rows they would make the reader map
   "solid" → RoadMarkType::Other — the fastest way to reintroduce xodr: unknown enum spellings are rewritten on save — lane type → "none", road-mark type → "solid", colour → "standard" #476.
2. Empty-string policies stay at the call site. value_of returns
   nullopt for unknown; the reader keeps its own Other/nullopt policy.
3. type_str preservation is untouched — the writer already prefers the
   verbatim spelling over the enum name.

PR-B — docs-site (≈ -175 lines)

• B1 Delete docs-site/scripts/report-external-links.mjs — 125 lines of
  hand-rolled fetch + concurrency + timeout for a report that always exits 0.
  lychee is already a CI dependency; a weekly non-blocking lychee job without
  --offline replaces it in ~15 lines of YAML.
• B2 Delete docs-site/scripts/theme-css.mjs — a JavaScript parser of C++
  source (indexOf + a QColor(0x..) regex over theme.cpp) that breaks on
  any reformat. The repo already solved this for editor/resources/help/help.css:
  generate in C++, commit the output, byte-gate it from a gtest. Reuse that.

Acceptance — byte-identical output, not green tests

Tests can pass while a round trip silently changes bytes; that is how #476
survived. A throwaway differential probe reads and rewrites all 126 tracked
.xodr files plus tests/esmini/*.xosc on main and on the branch, and
diffs the bytes. Zero differences is the gate, and the gate itself is
sabotage-checked (swap two spellings in the table; the probe must go red).

Declined by the maintainer, deliberately out of scope

• Cutting the Renderer interface (10 pure virtuals, one implementation,
  no test double). It is a documented architecture non-negotiable in the
  project guide; kept.
• Recompressing the 13 MB of docs/design/ui-revamp/ PNGs. Git history
  already carries the original bytes, so recompressing grows the repo net
  unless history is rewritten. Not worth it for design screenshots.
• point_segment_distance (identical in geometry/simplify.cpp and
  mesh/junction_corner_detail.cpp) — the two subsystems share no internal
  header, and creating one to save 13 lines is break-even.
• The PhaseSemantics read/write twin — one enum, six values, ~30 lines.
  Not worth a new header.

Net: ≈ -410 lines, 0 dependencies removed (every one of the 13 is
load-bearing; sharp is already stubbed for licence reasons).