docs(live-verify): fold in #637-#668, correct the entries the merge burst falsified

[vibechoom]

Closes #660

Scope

#660 was filed when the doc's coverage stopped at #634, expecting only #637/#639/#642 to fold in. Seven more PRs merged in the meantime, so this pass covers all ten: #637, #639, #642, #644, #645, #662, #663, #664, #666, #668.

The correction #660 asked for

The #630 entry told a verifier that modal::close_all tears down with window.close(). #637 changed that call to window.destroy() after #630 had already merged, so the entry named a call the code no longer makes — confirmed against the current modal.rs:1019 (panel.window.destroy();). Corrected the entry and added a note explaining the timeline (the entry was true when written, false the moment #637 landed).

Other falsified-expectation sweep

Per the task, I looked for other entries the merge burst falsified, particularly around the networkd startup path (#645 changed it from "go inert forever" to "retry unboundedly"). Found no existing entry describing the pre-#645 networkd startup-inertia behavior — it was never previously documented — so this is purely additive: added a new #645 bullet to the "Network panel (link status)" section, explicit that it's a different code path from the #610/#623/#634 backend-choice probe already documented there (the networkd seed_links first-ListLinks retry, not probe_link_backend's NM-vs-networkd choice). Verified no other entry in the doc references an inert-until-restart networkd expectation.

New entries added

• fix(overlays): destroy unrealized layer windows on teardown, not close() (#632) #637/fix(hytte-ui): destroy() the three remaining close()-teardown sites (#638) #639 — the close() → destroy() sweep across the five trollshell overlays and the three remaining hytte-ui sites (bar.rs, popup.rs). Carries fix(overlays): destroy unrealized layer windows on teardown, not close() (#632) #637's own live-verify gesture (glib::WeakRef / gtk::Window::toplevels() count across repeated kanshi hot-plug cycles for never-opened layer windows) verbatim, since it exists nowhere else.
• fix(overlays): release RefCell borrows before the GTK call that can re-enter them (#631) #644/fix(ui): release RefCell borrows across GTK calls outside overlays/ (#643) #663 — the RefCell-borrow-across-GTK-call sweep (31 sites total: fix(overlays): release RefCell borrows before the GTK call that can re-enter them (#631) #644's nine inside overlays/+modal.rs, fix(ui): release RefCell borrows across GTK calls outside overlays/ (#643) #663's twenty-two outside it, including the 8-panel-wide reactive_list.rs helper and the control-center's Plugins tab). Both use the same "nothing changes" honest pass condition fix(modal): drop the PANELS borrow before closing drawer windows (#627) #630's entry already established for this class of latent-hazard fix.
• fix(mpris): bind the Auto chip to selected_player, widen bare-string artist (#651) #664 — mpris Auto-chip fix + bare-string-artist widening. Flagged its two easy-to-get-wrong requirements per the task brief: needs two live MPRIS players (the switcher hides below 2), and the new CSS class needs a shell restart to apply.
• feat(claude-bridge): keyless loopback OpenAI-compatible shim over headless Claude Code (#584) #666 — new top-level "Claude bridge" section for the keyless loopback OpenAI shim; nobody has run this service live. Carries the PR's own flagged riskiest assumption — hive-claude classifying a failed --resume as the typed SessionNotFound — with its journal tell (does not match any session title in the stderr tail), since a drift there means every first turn silently 502s.
• fix(bus): log and back off a contested bus name instead of retrying silently at 4 Hz (#653) #668 — new top-level "D-Bus name ownership" section. States plainly that recovery going from ~250ms to up to 5 minutes after a squatter exits is the expected cost of the fix (traded for cutting RequestName from ~14,400/hour to 12/hour), not a regression to file.

Left alone

• fix(places): correct the shipped default station id (#641) #642 — no doc change beyond the coverage-line bump, matching the issue's own conclusion: write_default_config only writes places.toml when absent, so existing installs (including Annika's) are unaffected either way, and that's already handled by the Departures broken #641 triage comment + Control center places editor incomplete #640 tracking the editor-UI gap.
• The sites fix(overlays): release RefCell borrows before the GTK call that can re-enter them (#631) #644/fix(ui): release RefCell borrows across GTK calls outside overlays/ (#643) #663 themselves list as "deliberately not touched" (prompt.rs's abort_subscription, plain write-backs, etc.) — no live-visible behavior, nothing to verify.

Gates

• Diff is markdown-only: git diff --stat → 1 file, docs/live-verify.md.
• nix fmt run twice — 1 file reformatted the first pass (a prose reflow), 0 changed the second.
• Clippy/tests deliberately skipped: nix/package.nix's crane source filter excludes markdown entirely, so a markdown-only diff produces a bit-identical trollshell-workspace derivation — running nix build would burn a full workspace recompile for zero signal (the same crane-filter trap docs: correct claims in CLAUDE.md, CHOOM-INIT.md and the etc/ deployment READMEs (#658, #659) #662's own PR body documents). Not run, per task instructions.