fix(channel): clear reloading flag event-driven instead of fixed 200ms

[data-diego]

Important

This PR is fully vibe coded, I just preferred providing a solution instead of just ranting in an issue, feel free to close it and use it only as inspiration if its not up to the repo's standards.

Motivation

I'm porting a shell-script session picker to a television channel. It refreshes a tmux/sesh list every ~250ms and each row has a small spinner indicator for whether the Claude session attached to that window is waiting for input.

Under fzf (with reload-sync) the refresh is invisible: old rows stay on screen until new rows are ready, then they swap in one frame. Under tv <channel> --watch 0.3 the list flickers on every refresh — noticeable on its own, very visible on the spinner — because the blank frame lands right where the glyph would animate. Same source, same cadence, so this is a --watch rendering issue in tv.

Root cause

Two things compound in Channel::reload():

1. matcher.restart() drops the previous snapshot up-front, leaving the matcher empty for the whole source runtime.
2. Television::should_render is AND-gated on !channel.reloading(), a flag cleared by a fixed 200ms sleep. The UI is frozen for 200ms regardless of source speed — and if the source is slower than 200ms (e.g. fd . $HOME --type f, ~270ms), the flag gets re-set on the next watch tick before the previous load finishes, so the UI never unfreezes.

Fix: fzf-style atomic swap

reload() now builds a fresh staging matcher and feeds the new source into it. The live matcher keeps serving the previous snapshot — results(), find(), rendering, scrolling all stay responsive — until tick() swaps staging in as a single statement. The swap is purely event-driven: it fires when staging has items, or when the staging crawl task exits with no output.

Supporting pieces:

• find() is mirrored onto staging so the post-swap snapshot is already filtered to whatever the user typed during the reload window.
• SortStrategy::Custom holds a Box<dyn SortFn> that isn't Clone, so we carry a SortStrategyFactory closure that builds a fresh strategy per staging matcher. Frecency caches are Arc-cloned in.
• running() ignores staging. During a reload the user sees a stable list, so the UI's loading indicator should stay off. The initial load still reports running (no live results yet), and after a swap the former staging handle moves into crawl_handle so running() stays accurate while the new source drains.
• The reloading flag, the should_render gate, the reload_deadline field, and the RELOAD_RENDERING_DELAY constant all go away — the atomic swap replaces them.

Behavior change

A source that hangs without producing output no longer times out to an empty list after 200ms — it keeps showing stale data. This is strictly better than flickering to empty: the stale data is at least coherent, and --watch retries on its next tick anyway.

Repro

Before this PR, this flickers every 300ms on most terminals:

cat >/tmp/watch-flicker.toml <<'TOML'
[metadata]
name = "flicker-demo"

[source]
command = "seq 1 30 && date +%S.%N"
TOML

tv --cable-dir /tmp --config flicker-demo --watch 0.3

After this PR the list is stable; only the clock value (last line) changes in place. A source slower than 200ms (e.g. fd . $HOME --type f) also no longer flickers to empty.

Tests

• test_reload_swaps_when_items_arrive — swap fires as soon as the new source produces output.
• test_reload_swaps_when_empty_source_finishes — with a source that emits nothing (true), swap fires once the crawl task exits.
• test_reload_keeps_old_results_visible_until_swap — against a slow source (~80ms), result_count() stays at the pre-reload value for the full reload window, then flips to the new value after the swap.
• test_reload_does_not_swap_to_empty_while_source_hangs — against sleep 5, staging stays in flight well past any prior time cap; old entries remain visible.
• test_running_ignores_staging_but_not_initial_load — running() is true during the initial load and false during a staging-only reload.

cargo test --all --all-features, cargo clippy --all-targets --all-features -- -D warnings, and cargo fmt --check all pass locally on macOS.