0061 — The ingest gap was a collection, landing wherever the collector ran

Status: accepted · Date: 2026-08-30 · Amends: 0058

Context

0058 answered #480 and left one thing open:

ingest_total measures 2.17–2.26 ms where its own parts sum to 0.97–1.00 and the canonical harness measures the same chain at 1.109–1.134.

It named position within the round as the likeliest candidate — ingest_total is always first, so it absorbs whatever the previous round left the collector — labelled it a candidate, and said a phase-reordering run is what would settle it. This is that run.

What was measured

ingest-phases on 3d4fa2d, a hosted runner, .NET 10.0.11, 4 cores, workstation GC, three rounds of nine interleaved runs, load average 3.46 / 3.46 / 3.08. ingest_total_last calls the same Ingest method as ingest_total, last in the round instead of first, so a difference between them is position and nothing else.

	round 1	round 2	round 3
ingest_total, first	0.936 ms	0.949 ms	0.910 ms
its collections, gen0/1/2	0 / 0 / 0	3 / 3 / 3	0 / 0 / 0
ingest_total_last, last	1.308 ms	0.943 ms	1.255 ms
its collections, gen0/1/2	3 / 3 / 3	0 / 0 / 0	2 / 2 / 2

Canonical harness, same runs: 1.163 / 1.136 / 1.154 ms wall.

Decision

The gap is a garbage collection, and it lands on whichever phase the collector happens to run in. It is not position. 0058's candidate is refused by its own test.

Round 2 is what settles it: the two rows are 0.949 against 0.943 — indistinguishable — and it is the first that carries the three collections. Rounds 1 and 3 have the last one slower by 0.35–0.37 ms, and there the collections are on the last. The cost follows the collections, not the rank. A position explanation would have put the same row ahead in all three rounds.

And the row was never stable in the first place. Adding a second ingest to the round moved ingest_total from 0058's 2.17–2.26 ms to 0.910–0.949, without touching the ingest. With one ingest per round, that single row absorbed every collection the round's 15.36 MB allocations provoked; with two, the debt is split and neither carries all of it.

So 0058's unexplained gap is explained, and the number that raised it was an artefact of the table's own shape. At 0.910–0.949 ms the first-position ingest now agrees with both things 0058 said it disagreed with: the sum of its parts (0.97–1.00) and the canonical harness (1.136–1.163).

Nothing about 0058's substance changes. The ingest is still memcpy-bound, the reader's allocation still costs about 0.02 ms, and adopting is still worth exactly one copy. What changes is that the one row which did not fit that account now does.

What was refused

Shuffling the phases. Every ratio in this table rests on the phases being interleaved in a fixed order, so reordering per round would have made each round incomparable to the last in order to answer a question about one row. Duplicating one row at the other end of the same round is the same experiment with one variable instead of two.

Reading round 1 or round 3 alone. Either would have confirmed position handsomely — 1.308 against 0.936, 1.255 against 0.910 — and been wrong. Round 2 is the round that decides, and it decides against the hypothesis the run was built to test.

Chasing the collection out of the diagnostic. A GC.Collect between phases, or a larger allocation budget, would flatten the table and hide the thing this lot just learned: on a 15.36 MB block the collector's timing is worth 0.35 ms to whichever phase it lands in, which is a third of the ingest.

Consequences

• #480 is closed. Both halves are answered: 0058 for where the time goes, this for the row that did not add up.
• A single-call median on a 15 MB allocation is not a stable measurement here, and the phase table now shows why rather than only its result. Any future phase in this mode that allocates a block of this size should expect its median to carry a collection it did not cause.
• The canonical harness was right all along. Best-of-five over scaled iterations read 1.109–1.163 ms across both runs while the single-call median moved from 2.2 to 0.9 — which is an argument for what the published rows already use.
• What would change this decision is a run where the collections and the slower row come apart: a round with the collections on the first and the last still slower, or the reverse. Three rounds is what this rests on, and two of them agree by putting both together.