Render CI timing histogram through .dag (std.render Frame), not hand-rolled Rust boxes

[briansrls]

Redo of this PR's box-width work the principled way (the prior version hand-rolled
the boxes in Rust — a parallel re-implementation of std.render.Viewport +
extdeps/render/terminal::constrain_frame, flagged in review).

Problem

The timing-histogram boxes were built in Rust with fixed ~78-col borders under
~95-col content rows: too short in wide viewers (GitHub Actions), wrapping in
narrow ones (dashboard chat). The width was never modeled.

Fix — render through the .dag substrate (d5b3a85)

• extdeps/render/terminal.dag gains the missing primitive serialize_frame
  (Frame -> String, color-gated by RenderCapability via ansi_code + resolve_symbol).
• dsl/gunbc/ci_render.dag (new) builds each percentile box as a std.render.Frame,
  bordered + padded to a width parameter (the medium's Viewport.width) — every line
  uniformly that width. Single authority for box width.
• claim_executor evaluates render_percentile_box via run_in_context_with_args
  (the seam eval_spawn_width already uses) — a real consumer, green by execution. The
  seed supplies only measured percentiles + the host viewport width (COLUMNS, else 88,
  clamped); all layout/formatting lives in .dag.
• cli_run.rs drops the Rust box renderer + format_nanos, keeping only
  compute_histogram_data (the measurement the evaluator can't get otherwise).

§3 single-authority (review fixes)

• Consume std.render.repeat_string (deleted the local repeat_str fork).
• Consume std.unicode.string_display_width + std.width.truncate_text, and fix them:
  both were dormant-broken (chars yields Char = Int where unicode_scalar, but
  char_display_width called code_point(c) which wants a String, and truncate_text
  did concat(String, Char)). char_display_width now reads the scalar directly,
  truncate_text rebuilds via from_code_point — string_display_width works for the
  first time (correct CJK/combining width via its block tables).
• Box borders source glyphs from the registry (resolve_symbol + Tier) instead of
  hardcoded ┌─│┘ literals: 6 new Box* SymbolIds with single-column ASCII fallbacks,
  so a non-Unicode tier yields a +--+/| box automatically (no §5 mojibake fail-open).

Verified

Witness ci_render_histogram_width_test: per-line string_display_width == width at two
widths, exact Unicode 3-line oracle, ASCII-tier width holds, tier swaps glyphs.
cargo build / fmt --all --check / clippy --all-targets clean; full v1-compiler +
v1-compiler-tests suites green. Render failure is fail-visible but non-fatal.

Bundled: two pre-existing test breaks (gate-unblocking, test-only)

This PR touches v1-compiler Rust, which triggers the full test suite (main's CI
green only via affected-test scoping). That surfaced two failures pre-existing on
origin/main (both verified on a clean #5818 checkout), unrelated to this diff:

• 4ca65ea — contracts_sidecar_wired_into_emit_scope: loose .find() glob matched the
  split-out anthropic_errors.rs instead of the declaring anthropic.rs (the contract
  emits correctly; verified). Tightened to the declaring module's file.
• a21996b8 — string_family_cast Rust test: Remove hollow Url=String alias (§3); ground RestTransportConfig.base_url as Uri #5818 removed type Url and updated the
  .dag witnesses but missed this .rs test's inline Url. Mirrored Remove hollow Url=String alias (§3); ground RestTransportConfig.base_url as Uri #5818's edits.

Deferred to a follow-up

The [file] read host-trace spam-collapse — done faithfully it routes the trace sites
through gunbc.output_policy.channel_decision (evaluated once at startup, not a Rust mirror).

[gunbai-bot]

Addressed all four review findings in d5b3a85 (the WIP push the bot reviewed was an intermediate auto-commit; force-pushed clean now):

1. zz_temp_diag_anthropic_emit — removed; it was a local debug probe that shouldn't have landed.
2. repeat_str — deleted, now imports std.render.repeat_string.
3. text_columns — deleted, now uses std.unicode.string_display_width. Note: that authority was dormant-broken (chars yields Char = Int where unicode_scalar, but char_display_width called code_point(c) which wants a String; truncate_text did concat(String, Char)). So I fixed the authority rather than fork it — char_display_width reads the scalar directly, truncate_text rebuilds via from_code_point. string_display_width now works for the first time (correct CJK/combining width via its block tables), which also resolves your §5 point.
4. Box glyphs — now sourced from the registry via extdeps/render/glyphs.resolve_symbol with a Tier (6 new Box* SymbolIds with single-column ASCII fallbacks). A non-Unicode tier yields a +--+/| box automatically — no hardcoded literals, no mojibake fail-open. Witness now also checks ASCII-tier width + that the tier swaps glyphs.

Separately, 4ca65ea fixes a pre-existing contracts_sidecar_wired_into_emit_scope failure (verified on a clean #5818 checkout — surfaced only because this PR touches v1-compiler Rust and triggers the full suite). Test-glob bug from the anthropic module split; the contract emits correctly.

— sent from bright-ibex-44

[gunbai-bot]

Addressed the blocking unit-modeling finding in 04751847:

• Histogram durations now land on std.measure.Nanosecond (Measure<Time, Nano, Nat>), consumed via nanosecond(count:) / nanosecond_count(n:) — no local unit type, no flat scalar. format_duration / timing_cell / percentile_row / percentile_box_frame all take Nanosecond.
• The seed↔.dag boundary stays Int (as you noted is fine — claim_executor passes raw measured nanos via clamp_nanos_to_i64); percentiles_box_frame_ns wraps those into Nanosecond immediately, so the unit-bearing type is the authority everywhere downstream.

Re the non-blocking point #2 (the std.unicode/std.width fixes): they're called out in the d5b3a85 commit message ("fix dormant std display-width") with the Char = Int rationale, and the new witness now exercises that path (ASCII-tier width + tier-swap both run through string_display_width/truncate_text). Happy to split them to a standalone PR if you'd prefer cleaner bisection — let me know.

— sent from bright-ibex-44