[sotashimozono]

The claim was too strong

A revisit costs zero remote round trips; conditional GET is answered inside localhost.

The second half is true and verified. The first half is true inside the freshness window and cannot be true outside it: once a listing is not trusted, "is this still the current version?" cannot be answered without asking.

Measured against a real host over SSH:

	revisit
straight after the load	0
~3 s later	10

That is the window doing its job. The README now says which window the claim holds in, and what a revisit costs outside it.

The e2e now checks it over the real transport

a_revisit_costs_no_remote_round_trips holds a FakeRemote and a three-byte fixture. It revisits instantly, so it never meets the freshness window — it would pass with any TTL, including none.

The new check drives a real browser against whatever host the run is pointed at, reads the round-trip count off GET /_control/hosts, and asserts a revisit adds nothing. Neutered against DEFAULT_TTL = Duration::ZERO:

ok    the first visit costs the remote something
FAIL  and a revisit inside the freshness window costs it nothing
      a revisit went to the remote, so it was not answered from cache
      85 !== 69

The unit test passes unchanged under that same neutering.

It navigates away and back rather than calling reload(). A reload tells the browser to re-fetch, and measuring it would not be measuring this claim.

What this came out of

An investigation into what looked like a much worse bug — "a revisit costs 13–30 round trips on every real site". It was not. Written up in full in #49, but in short:

• Driven directly, the daemon's cache is exact: repeat read 0 round trips, range 0, conditional 0 with a 304. Through the alias Host and through the loopback fallback alike.
• page.reload() is not a revisit.
• "0 conditional requests" was false — request.headers() is the pre-flight snapshot and misses headers the browser adds. The 304s in the responses prove revalidation was happening.
• 342 of 412 browser requests on one page were Chrome's own PDF viewer loading its internals, not requests to our origin at all.

I also built the listing TTL as an idle window rather than an absolute one, and reverted it: it passed neutered tests but made no measurable difference, and moved the Documenter page in the wrong direction. Shipping a public API change on an argument the measurement does not support is not how this repository works.

Verified

cargo fmt --check, clippy -D warnings, 177 tests, and the full e2e against a real host: all checks pass.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[sotashimozono]

Pushed a fix, and the reason is worth writing down

The check I added here failed in CI while passing against a real host. Twice over, the instrument was part of what it measured.

It sampled the counter off /_control/hosts, which runs ssh -G once per configured host. Taken either side of a page load, that call takes long enough to expire the listings being measured. The count now comes off hello — the cheap route, two milliseconds — and reporting it there is part of this change.

With a fast counter it then failed everywhere, because the slow call had been acting as a settling delay. Without it, the first load's tail landed in the second load's bucket.

And underneath both: it asserted something nobody had established. Invariant 2 is a claim about a request. "A whole browser page costs nothing the second time" is a larger claim and it is false — against a real host this fixture costs four round trips on a browser revisit, because a page fetches things no HTML scan can see, and the freshness window is a wall clock a test has to race.

So the check now tests what invariant 2 actually says, over the real transport, where it is exact:

ok    the first read of a file costs the remote something
ok    reading it again costs nothing
ok    and a range out of it costs nothing
ok    the browser's own validator is answered here, not there

Neutered against DEFAULT_TTL = Duration::ZERO, the last three fail (4 !== 0). The unit tests pass unchanged under the same neutering, which is why these are worth having as well.

This is the third time today that a measurement here turned out to perturb its subject — the other two are in #49. The pattern is worth naming: a counter you cannot read without disturbing is not a counter.