fix(ui): strip ATX heading markers from quote chip display

[cat0825]

Fixes #2213

Problem

A quote chip can carry markdown source verbatim (selection capture keeps the
raw text), so a heading line like ### Description renders in the chip as
cramped, overlapping literal hashes — it reads as a blob instead of three
distinct #.

The markdown renderer is not at fault: ### mid-line is not a heading marker
per CommonMark (/^(#{1,6}) +(.*)/ anchors at line start), so it correctly
renders as plain text. The issue is presentation of that literal text in the
chip.

Fix

Add stripQuoteHeadingMarkers — display-only cleanup that removes a leading
ATX heading marker (#{1,6} ) from the quote text shown in the chip:

• ### Description → Description
• ###Description (no space) → unchanged (not a heading)
• C# code, #hashtag, mid-text # runs → unchanged (treated as content)

Presentation only: formatTextWithInlineRefs still sends quote.text
verbatim to the model, matching the chip's existing contract.

Tests

packages/ui/src/__tests__/quote-ref-chip.test.ts — 3 tests covering marker
stripping (1–6 hashes, multi-space/tab), and non-heading preservation.

Run:

• npm --workspace @maka/ui run build (tsc) — passes
• node --test dist/__tests__/quote-ref-chip.test.js — 3/3 pass
• Full @maka/ui suite: 324/326 pass; the 2 failures
  (astryx-form-controls-localization) are pre-existing and unrelated
  (Astryx TextInput "Optional" marker localization, not touched by this PR).

[Astro-Han]

Clean minimal fix — the regex is correct against CommonMark (####### x correctly left alone), the model still gets quote.text verbatim (model-history.ts:906), and the pure-function tests are solid. One gap worth closing before merge, plus small notes:

• P1 — the pre-send drawer shows the same bug: quotes created via text selection carry no label (chat-view.tsx:603-607 → use-app-shell-composer-quotes.ts:27-34), so the pending-quotes Token at composer.tsx:1234 falls back to raw quote.text.slice(0, 48) — ### Description still renders as cramped hashes in the drawer, one step before the chip you fixed. One line: strip only the fallback branch — label={quote.label?.trim() || stripQuoteHeadingMarkers(quote.text.slice(0, 48)) || copy.pastedQuoteLabel} (covers the quote-companion composer too, same shared component).
• P2 — test gap: add assert.equal(stripQuoteHeadingMarkers('####### x'), '####### x') — the one case where a naive /^#+ / would diverge.
• P3 — docstring: "the only marker that can survive is #{1,6} at the start" overstates — selection normalization (use-message-selection-quote.ts:46) also preserves - , > , 1. prefixes; they're just out of scope.

P1 is a few lines (and pre-send, users see the very symptom #2213 describes); everything else can be follow-ups. Nice work otherwise! 🚢

[cat0825]

Addressed all three review points on fix/2213-quote-heading-markers (plus synced upstream main, fixing the composer.tsx import conflict with #2230):

• P1: composer.tsx pending-quotes Token label fallback now routes through stripQuoteHeadingMarkers — quote.label?.trim() || stripQuoteHeadingMarkers(quote.text.slice(0, 48)) || copy.pastedQuoteLabel. Covers the quote-companion composer too (shared component).
• P2: added ####### x and ######x boundary cases to the pure-function tests.
• P3: softened the docstring — other line prefixes (- , > , 1. ) survive selection normalization; only the ATX marker is stripped, by design.

Verification: @maka/core + @maka/ui tsc builds pass, quote-ref-chip.test.js 3/3, biome clean.

[Astro-Han]

Thanks for this — the fix is clean and I verified both display surfaces: the pre-send token in the composer (label-less selection quotes) and the sent chip's displayText path both strip the marker, while the model still receives quote.text verbatim (display-only contract intact). The branch's net diff against upstream main is exactly the 3 files — the extra commits in history are the upstream merge (the #2237 perf commit is already in main; nothing unrelated lands). Merge is clean, and the 14 assertions genuinely pin the regex (they'd fail pre-fix).

Three optional follow-ups, none blocking:

• The doc comment says CommonMark allows "space (or end of line)" after the marker, but the regex only handles space/tab — a quote that is exactly ### survives unstripped (and stays truthy, so the fallback label never fires). (?:[ \t]+|$) would match the documented contract.
• bug(ui): markdown heading markers (###) shown verbatim in quote chips render as cramped literal text #2213's screenshot repro (…/issue.md ### Description, marker mid-string) is pinned as unchanged by the tests — CommonMark-correct, but worth confirming against the live app whether the reported visual actually had the marker at line start (if not, the issue needs a note on what exactly this PR fixes).
• Test coverage is pure-function only: the sent-chip render path, the composer fallback chain, and unicode-after-marker (### 中文标题) are untested.

Happy to approve — these can be tracked as follow-ups.