Refuse to report an ETA while every lane is still warming up

[EHxuban11]

Third in the series after #7 (wrong numbers) and #8 (incomplete artifacts). This one is wrong estimates.

The bug

Three campaigns, three badly wrong headline ETAs, all the same artifact:

campaign	reported	reality
ec-s, cache filling, 0 epochs done	est. remaining 64.9 h	unknown, but not that
rfdetr-n, fresh dataset, 0 epochs done	90 h	unknown
yolox-tiny (earlier)	16.3 h	actual campaign 4.4 h

A run's opening epochs carry the image-cache fill, allocator warmup and cuDNN autotuning. wall / epochs_done there is not a per-epoch cost, it is a startup cost being annualised across 100 epochs.

observations() already knew this — it excludes running datasets below MIN_EPOCHS_OBSERVED from the cost-model fit. But the per-lane remaining-time loop called _epoch_seconds(record) directly with no guard, so the contamination came back in through the other door and dominated the makespan simulation.

What changes

• A lane contributes its own per-epoch rate only once past MIN_EPOCHS_OBSERVED; before that it falls back to the fitted model, which was built from vetted points.
• When lanes are running but none has cleared warmup and nothing has finished, return available: False plus a reason, instead of a number.
• Dashboard renders ETA not available yet (warmup/cache fill) with the reason as a tooltip.

The distinction that matters

A campaign that hasn't started has no misleading evidence, just no evidence — that keeps the long-standing p50 = 0.0 / "no observations yet" behaviour. Only the case where lanes are actively running but all still warming returns unavailable.

I initially conflated those two and broke test_eta_reports_its_own_method_when_it_knows_nothing. Narrowing the guard was the right fix rather than weakening that assertion.

Check

tests/test_eta_refuses_to_guess.py (new, 5 tests): the real ec-s situation returns unavailable; zero-epochs returns unavailable; it becomes available once one lane clears warmup; available as soon as any dataset finishes; and a warming lane with a huge wall does not drive the estimate.

Full harness suite: 158 passed.

Not verified

• MIN_EPOCHS_OBSERVED = 5 is inherited, not tuned here. For a ~250 s epoch that is ~20 minutes of "no ETA" at campaign start, which I think is the right trade against a 10x-wrong number, but it is a judgement call.
• The p90 band and predicted_total_seconds are unchanged; only the availability gate and the per-lane rate are touched.