Reported symptom: "!link arrived only on the second try". Cause was the throttle
itself -- one global 10 s window per channel, so a second person asking inside it
got silence. That reads as a broken bot rather than "wait ten seconds", and it
happened twice while testing.
Now each sender has its own window, keyed on the "<name>: " prefix, over an
8-slot LRU. Spoofable by design and that is fine: this is politeness throttling,
and a channel message carries no identity beyond that prefix anyway.
The airtime guarantee that the global window used to provide is kept explicitly
as BOT_MIN_GAP_MS (2.5 s floor between any two replies, whoever asked), so N
senders can no longer produce N simultaneous transmissions. Without that, going
per-sender would have quietly traded politeness for responsiveness.
`bot` now reports both numbers plus a throttled counter, so a suppressed reply is
visible instead of looking like a dead node.
Bench-verified on the discriminating case, which is the whole point: sender A
asking twice in 3.5 s gets one answer, while sender B asking INSIDE A's window
gets answered -- 3/3.
Note on how that was measured, because it bit me: reading the console counter
costs ~2.4 s (the CDC needs a settle delay), so a probe pair with a counter read
between them is ~15 s apart and the 10 s window has legitimately expired. The
first run of this test "failed" for exactly that reason. Read counters only
before and after a pair, never inside it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>