fix: connector_detached now skips fully dash-hidden connector shafts

[miga-heygen]

Summary

connector_detached (the layout-audit finding for SVG connector paths whose rendered endpoints don't attach to any anchor) evaluated a path's geometry unconditionally — it had no gate for stroke-dashoffset/stroke-dasharray draw-on visibility. A connector fully hidden behind an un-advanced "draw the line in" dash animation still fired the finding as if it were rendered and visible. The sibling check, connector_orphan, already skips fully dash-hidden shafts via a shaftDashHidden(path) helper — this adds the same gate to connectorDetachmentIssues, in the same position.

Deliberately did not also add the shaftIsPainted check that connector_orphan uses (which additionally covers opacity/display:none/visibility:hidden) — that's a real, separate gap but out of scope for this fix; noted below as a follow-up candidate.

Test plan

• Added a regression test: a connector fully hidden via strokeDasharray/strokeDashoffset no longer fires connector_detached, using the exact fixture (now shared via extracted foreignFrameDom/foreignFrameRects/foreignFrameStyles consts, mirroring this file's existing orphanDom convention) as the adjacent positive-control test that proves the same fixture does fire when not dash-hidden — isolating the single variable.
• Verified via a real-Chromium (puppeteer) revert-and-restore harness against the actual browser script (this file has no build step and can't run under this sandbox's happy-dom test runner): confirmed pre-fix, a dash-hidden connector fired the identical finding as a fully-visible control; confirmed post-fix, the dash-hidden connector reports zero findings while the control still fires.
• bunx tsc --noEmit -p packages/cli, oxlint, oxfmt --check clean on both touched files.
• Full packages/cli test suite: 3040 passed, 3 skipped; the only 2 failures are in files unrelated to this change and fail identically on an unmodified checkout (happy-dom collection error, pre-existing sandbox limitation — this PR's own regression test lives in one of the files affected by that same pre-existing limitation and could not be executed locally, only via the real-Chromium harness above).
• Adversarial review also flagged: shaftDashHidden and pathUserEndpoints each independently call path.getTotalLength(), a redundant SVG geometry query — but this exact double-call pattern already exists today in the sibling connector_orphan check this PR intentionally mirrors, so it's not a new inefficiency.