[tony]

Summary

• Fix unrelated Java links on Clojure argument names.
• Preserve hyphenated var names and native declaration links.
• Keep authored references and Java type hints resolvable.
• Rebaseline only Clojure’s cross-reference minimum after removing false signature links.

In (send! connection argv timeout), timeout linked to Java's CommandRequest.timeout. The tokenizer also split join-pane! and one-or-none into fragments and linked those fragments to unrelated methods. Native Clojure call forms now identify the public var and its local bindings before generic symbol lookup.

The renderer regression exercises all 85 public functions and 98 arities with the actual Java parent inventory. Other language branches retain their existing behavior.

Test plan

• Reproduce the accidental timeout link with the original renderer.
• Pass 231 focused Clojure, JVM, scope, type-parameter and inline-link tests.
• Pass the actual Astro signature renderer across every public Clojure arity.
• Reject the original renderer, restore the fix, and pass the renderer regression.
• Pass API-model lint, type checks and git diff --check.
• Build the fix combined with PR Fix five ports' docs builds and tidy the port landing page #296 through the actual Clojure publisher inputs in 113.05 seconds.
• Verify 3,382 provenance files, 14,175 internal links/anchors and 3,440 assets; reject Java type links inside native call forms.
• Check four Clojure pages at 280, 390, 768 and 1440 pixels in Chromium; exercise JVM grouping, keyboard focus and two deliberate rejection controls.
• Audit source-matched publisher output: 53 false signature links removed, one correct namespace link added, and all 311 references outside signatures unchanged. The checker passes at 643 and rejects 644; other ports’ minimums are unchanged.
• Rebase both implementation/test patches unchanged onto docs main 856211d85; pass the focused model, renderer, type and ratchet checks.
• Complete the broad outer gate. On the rebased branch, pnpm test failed after 40.96 seconds: 1,454 model tests passed and 14 skipped, but the full-corpus resolver test exceeded its unchanged five-second timeout (11.11 seconds under parallel load). Pending checks were stopped. Earlier isolated success is not a broad pass; no timeout or loop budget is increased.
• Hosted CI.

This draft targets current docs main. Coordinate landing with the shared docs release; the no-preview label holds preview publication.

Local publisher qualification uses an unpublished integration commit, 7cf7d80f20cc8dea8437f51342b2c794fb91cde8, combining PR #296 at 0149aad1 with this fix at 2d326686e. It renders 124 Clojure pages, including 92 reference pages, from clean source records. The production type-link checker reports six linked types and five plain ns keyword tokens. All native call forms are free of unrelated Java links. This local result leaves full shared CI and publication of this renderer fix pending. The r19 baseline documentation is already publicly published; it does not contain this follow-up fix.

The browser check of that same build passes 16 page/width combinations in 6.52 seconds, including the real shared native-header geometry checks. The picker places Clojure with Java, Kotlin and Scala; filtering, Escape and focus return work. Covered controls and an injected false Java link are rejected, then restored. Chromium closes with no remaining owned process. Firefox failed at startup and supplies no rendering evidence. External requests are blocked or replaced by the shared checker; other ports and public deployment remain outside this check.

The original hosted run failed at the Clojure cross-reference minimum: 643
against 673. The before/after output measures 695 → 643 because 39 Java
links and 14 own-symbol/partial-prefix links inside signatures disappear,
while the corrected libtmux.core-async namespace contributes one valid
link. All 311 references outside signatures are identical. Set the Clojure
minimum to the measured 643; retain the shared checker and every other
port’s record. A deliberate one-count shortfall still fails.

The rebase preserves both original patches exactly. The linker is byte-for-byte
identical to the already-qualified local publisher input; the new commit only
updates the measured Clojure minimum. The outer failure above remains open,
and this PR remains a draft while its new hosted CI run is pending.