docs(unic-archon-dlc): record the sixth check as shipped, and what building it taught

[orioltf]

Why

WI 43028 merged this morning as PR 5854 (f77ee51 on DXP-DesignSystem's develop), and the run 2 findings register still described it as work to do. That row is the one #457 reads to know whether run 3's last precondition is met, so leaving it stale would leave the frontier wrong.

The bigger reason is the new F5. Building the sixth check produced more than the check. Three unic-dlc-pr-review iterations found twenty-three things, and the ones worth keeping are not about that check — they are about a green that means nothing, found in four places in one day, three of them inside the instrument run 3 will be read through. None of it was written down anywhere durable; it existed in one session's window and in ticket comments.

What changed

One file, docs/research/2026-08-31-dlc-run-2-findings.md.

• F1's "a sixth check" row now records the artefact: merged PR, commit, what shipped, and that countStories became the one reader of the tag arithmetic so 43022 and 43028 cannot drift apart.
• F4 stops saying "the five checks". Six ship, and the F1 table is the count rather than a sentence beside it.
• F5 is new, six rows:
  1. A mechanism can read green by construction. 43028's check passed 12 tests, four of them failure fixtures, and still returned [] on the barrel shape develop gets the day the first component lands.
  2. So every check gets a positive control — obligation universal, technique per-check. 43020 is the counter-example, and four of the six controls are still owed before run 3's result is read.
  3. With no CI in the Consumer, the PR description is the only record that a check ran. Quote the runner's output rather than asserting it.
  4. /pr-review iteration 3 produced eight findings, wrote none, exited 0 — filed as bug(unic-archon-dlc): /pr-review produced eight findings, wrote none, and exited 0 #465.
  5. Four of the Consumer's eight declared commands cannot fail.
  6. estimations moved off → both between the runs, retiring run 2's cost and halt figures.

Rows 5 and 6 are what the seal marks not-comparable rather than scores.

The one judgement worth reviewing

F5 row 2 exists because I got it wrong first. My original wording on #457 generalised 43028's control technique to all six checks; 43020 breaks it, because uiReactStyleSheets() throws on empty rather than returning []. The DS-43028 session made the same move inside the README paragraph it added to stop over-claims. Both instances are in the row on purpose — the rule is worth more than either:

A quantifier over a set of mechanisms is an unverified claim per member until each member is named.

If that reads as too much narrative for a findings register, say so and it compresses to the rule alone.

Verification

prettier --check clean on the changed file (it re-padded the F1 table, which is why the diff is larger than the prose added). No package touched, no version bump, no changelog entry — this is a research document.