Record 2026-08-26 in SESSION-NOTES.md

[reissjason]

Notes only — one file. Updated after the CRT implementations landed, so it records six implemented CRs rather than seven awaiting review.

What the day was

Prototype (#36, #37, #38): three dangling doc links, two of which had never resolved because their targets exist only on the unmerged internal-docs root commit; CR-2026-035, where fixing one stale sentence exposed unreproducible benchmark figures and ten wrong cells in the C feature matrix; CR-2026-036, the metadata enrichment block.

LA-PS: CR-2026-013 and CR-2026-014 had been sitting uncommitted on a six-day-old branch while this repo had already shipped both — git grep expected_warnings HEAD returned nothing on any branch while nine files here depended on it. Then CR-2026-036, PS-309 to PS-325.

LA-IL — six CRs implemented, IL-132 to IL-162. CRT-2026-003 through 007 submitted and implemented the same day, then CRT-2026-008 out of implementing them: eight identifiers were declared in bold inside the mapping guides while the clause introducing those guides declared them non-normative. Kept and scoped rather than removed, using the per-format claim IL-119 already defines.

Numbers

Python	2706 → 2819
Corpus	1250 → 1253
Encode round-trip	1163 → 1166
C harness	488 → 491

Flagged rather than fixed

The floors were not raised by CR-2026-036. Decode is 1242 against a floor of 1239; reference encode is 1166 against FLOOR_TOTAL = 1163. Three vectors of headroom in each that nothing locks in — an edit that starts losing them would pass. Left as your decision.

Six mistakes, all one shape

Measuring against a model instead of the artifact:

1. The cross-reference fix was wrong by one, and my own checker agreed — both assumed one included file equals one numbered section. Only the PDF's table of contents disagreed.
2. I then verified that correction against a PDF five weeks stale.
3. Two guard simulations reported success against guards that were working — neither sed created the condition claimed.
4. A table cell with an unescaped underscore broke a PDF build with ! Missing $ inserted. and no line number, which pandoc -t latex cannot reproduce.
5. I concluded seven identifiers had been withdrawn because they were absent from spec/. They were in the companion guides, which that scan could not see. An absence is evidence about where you looked.
6. I allocated four identifiers over a submitted CR's proposals, by taking the next number after the highest in the specification text — the one place a pending proposal is guaranteed not to be.

The write-up treats the correction of #5 as the more useful entry than the record it corrected, including that the section now carries a machine-readable line, because the prose parser I wrote first read the correction and counted the seven it says are not withdrawn.

Where it says to pick up

CRT-2026-002, now unblocked with upstream evidence: the events refactor has landed in eclipse-thingweb, whose LoRaWAN converter models uplinks as events and rejects a Thing Description declaring properties. Our wot-td.md still describes output that converter refuses.

CRP-2026-001 gains a pointer to MultiTech-Systems/edgex-device-lorawan — live-verified uplink and downlink against an MTCAP3, §4b quality working — which did not exist when that CR was written.

Note for future editors

tests/test_cr_2026_035_benchmark_flagged_claim.py reads this file and asserts the corrected transform conclusion is still present. It is not free prose. That test is green (35 passed).

Verified

• pytest tests/ -q → 2819 passed / 4 skipped
• make docs-index-check → up to date

🤖 Generated with Claude Code