It refused to count until a block arrived while the page was open, on the
grounds that the block the snapshot hands over on load may be a minute old and
starting it at zero would lie. That was honest and completely useless: on
mainnet the block that triggers the counter is the same block that puts the
first row in the console, so it became visible and unnecessary in the same
tick. Reported from the live site after a refresh — the placeholder was still
there, because the counter had no window in which to exist.
It now seeds from the snapshot's block. The ticking is still
`performance.now()`, which cannot be skewed or stepped; only the seed can be
inexact, and only for that first block, whose age nothing but the server's
`observed_at` can attest. A seed outside a plausible range is refused outright
and the plain text stands — no counter beats a wrong one.
Measured against the live API: the counter appears within 800ms of load
reading "26,378 ms since #43,588", climbs, and gives way to rows about two
seconds later, which is how long the console is honestly empty.
Also recorded, because it wasted a verification: `/v1/healthz`'s `commit` is
stamped into the API binary, so a frontend-only push leaves it on the previous
sha *forever*. A watcher polling it for the sha just pushed waits for something
that will never arrive and reports "still deploying" over a site that finished
minutes ago. Compare the served bundle name to the local build instead — Vite
hashes it by content, so equal names mean byte-identical.
Refs #17
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
rob thijssen
67e7b2ec24
