[ako]

Refs #2160 — the reconcile, the last of the three parts of the settled page re-run.

This is in the identity code (ADR-0008), so the change is deliberately narrow and the old path is kept for comparison.

What was slow

canon.Reconcile was 23 s of a 77 s settled re-run of 600 pages, about 38 ms for each rebuilt page compared with its stored unit.

Reconcile is a row of steps from bytes to bytes, and each parsed both documents for itself: twelve whole-document parses for one write.

Step	Parses before	After
CarryTranslations	2	2
TransplantIDs: pairing	2	1, plus the shared stored parse
TransplantIDs: checking the substitution (ids before and after)	2	0
CarryPersistentIDs	2	0 for a document without the property
CarryUnexpressed	2	2
Equal	2	1, plus the shared stored parse

The change

• One parse of the stored document for the steps that only read it: the transplant's pairing and the final comparison (parseStored, transplantIDs, equalToStored).
• The substitution check reads ids off the raw bytes (elementIDSetOf) instead of parsing the document before and after it patches.
• The PersistentId carry is skipped for a document that does not contain the bytes of a binary PersistentId element. Only a workflow's elements have one, and a pair needs it on the new side.

What was deliberately not done

Parsing each document once and handing the trees to every step. The translation carry and the unexpressed carry copy parts of the stored tree into the new one, so a shared stored tree could be changed under the steps after them, and the new document changes between steps. Those two carries keep a parse of their own.

Evidence

• Byte for byte against the old Reconcile, kept verbatim as reconcileInPasses: every PedApp unit against itself, against a rebuild with new ids throughout, against rebuilds with properties left out, and the reverse, with and without ContentsOwnTranslations — 2,214 comparisons, same bytes and same verdict. The test asserts it covered unchanged and changed verdicts and reconciles that rewrote the contents.
• The workflow path: the fixture has no workflow, so a workflow-shaped pair is compared separately and must come back unchanged with the stored PersistentId.
• elementIDSetOf against parsing on every unit, as stored, rebuilt and thinned.
• Controls, each run and seen to fail: not reading ids under arrays fails the id test (4 ids read where parsing reads 801); always skipping the PersistentId carry fails three tests. A first attempt at the second control did not compile and printed nothing, and was redone.
• The existing canon tests pass unchanged, as do the executor's reconcile, idempotence, translation, GUID and workflow tests.
• End to end: a script that changes all 600 pages, run on two copies with the previous binary and this one, leaves 1,719 units that are byte-identical between the two; 600 of them differ from before the run. Script output is identical.

600 pages	Before	After
Settled re-run	74.3 s	61.6 s
Run that changes every page	72.1 s	65.6 s

Same machine, interleaved, one run each. The microflow and entity re-runs are unchanged within noise (6.9 s / 7.0 s and 5.7 s / 5.5 s).

Not done

• The remaining seven parses.
• The new comparison test takes about 27 s.

---

Generated by Claude Code