wayfinder: run 3 — do mechanisms carry the knowledge where documents did not? #456

Destination

Run 3 of the /specs → /tickets → /build chain has been dispatched against sealed predictions, read on both legs, and written up. Run 2 showed that documents do not reach the implementer. Run 3 asks whether mechanisms do. The map is done when the answer is on record — not when the run finishes.

Notes

Domain. The plugin is unic-archon-dlc in this repository; the Consumer is DXP-DesignSystem on Azure DevOps. The parent effort is #373 — read its body for the destination this one serves.

This map carries execution, deliberately. /wayfinder plans by default and its tickets are decisions; an effort overrides that in these Notes, and this one does. Most of what gates run 3 is work — an install fix, a Box refresh, a deletion, a parking move — and splitting the doing from the deciding would leave the map unable to say what is left.

The class rule, which is why this map exists at all. Every change queued before a run is one of three things:

A second change is a confound only when its effect cannot be told apart from the first's. That is why #441 landed before run 3 while looking like a confound: the checks test properties a PRD cannot mask, so both legs are separately observable in one run.

Half the treatment is on Azure DevOps, and no GitHub edge can reach it. ADO 43028 gates this map and cannot be a child or a blocked by here. It is named in the dispatch ticket's body and in the table below. A wayfinder group shrinks the prose residue; it never removes it. Two other cross-tracker edges already survive only as prose (#446 ← ADO 43018, #449 ← ADO 42998).

The Azure DevOps mirror, so the client sees this work. The client has no GitHub access, the daily runs off ADO, and only User Stories carry Story Points there. So each ticket on this map that is real work has an ADO User Story under Feature 42975, carrying title, a link to its GitHub issue, points and state — never a copy of the body. Nothing is duplicated, so nothing can drift.

Standing constraint. No dependency upgrade until phase 1 ships its first stable release. Archon is held at 0.7.0 on purpose. Nothing here proposes an upgrade as a fix.

Ticket shape, decided 2026-09-04 by the maintainer, applies from the next grilling on. A grilling closes its decision ticket, the /wayfinder way: the decision ticket gets the Decisions-so-far line and closes when the criteria are approved, and a separate implementation ticket carries the build, linking to the decision ticket's criteria rather than copying them. Each gets its own ADO User Story: the decision ticket's points are the grilling, the implementation ticket's are the build, both estimated when their shape is known. #430 and #439 predate this and stay single tickets with one mirror each.

Skills. /grill-with-docs for a ticket whose criteria are not settled — typed by the maintainer, never /grilling alone, which has nowhere for a decision to land. /writing-for-agents for anything an agent reads. /code-review before any pull request, and its four reads run twice: once on the criteria, once on the diff.

Decisions so far

(what was decided before this map was charted is in #373 § Decisions so far)

• A review node produced thirteen findings and recorded them nowhere
  — closed as not a defect, 2026-09-03. The premise did not survive a direct read: unic-dlc-build.yaml's
  report node folds the precheck findings in verbatim, and run 2's open-pr committed them in
  workflows/profile-card/report.md § 4 on the PR branch (c66eac9). An empty $ARTIFACTS_DIR — empty by
  design, evidence withheld — was read as "no file carried the output". What the grilling found instead,
  three /qa nodes whose prose nothing reads or keeps, is #458,
  ruled outside run 3. Register row 35 corrected. Dispatch run 3, read both legs against sealed predictions, and write it up #457 loses one blocker.
• The config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all
• Mechanisms reach the implementer where documents did not; the review Box's variance becomes false fixed state on the pull request — run 3, scored 2026-09-16 in the register § "The sealed prediction, scored — run 3" and § G; run 4's set is table G plus bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439, re-grilled first
  — merged 2026-09-05 as f9886f6 (PR #461), unic-archon-dlc 0.28.0.
  The config gains sdlc_needs, nine nullable keys naming needs and never tools; every Box installs once at bootstrap
  and says whether it did; every command outcome is pass | fail | unresolved, unresolved is never green, test is the
  floor at /build's evidence and /qa's merge; the unresolved needs land in a durable block (report.md, the posted
  review summary, a new qa-checks.md); /qa gains a test node it never had. Proven by one /qa run in
  DXP-DesignSystem on a config with no block: nine nulls, nothing fabricated, and qa-checks.md naming run 2's false
  pass unprompted. The positive path (install, test, pass, certify) has not run yet; it needs the maintainer's Consumer
  steps after the 0.28.0 tag. Instrument, so Dispatch run 3, read both legs against sealed predictions, and write it up #457 loses its last blocker and run 3 is dispatchable. ADR-0037.

Not yet specified

• How the result gets read. SETTLED 2026-09-15. The seal's shape is ruled: one arm per mechanism predicting the defect and the firing separately, a second arm of defects no mechanism guards, scored by the orchestrator against the run's diff and the check output, written into the register in section "The sealed prediction, scored". Full text on #457 comment 5683363982.
• What run 4 asks. ANSWERED 2026-09-16 by run 3. Mechanisms did not fail, so the planned-for branch did not open; run 4 tests the reviewer (register table G rows G4–G8) and bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439, re-grilled. Original text follows for the record. What run 4 asked. It depends entirely on run 3's answer, and the interesting branch is the one nobody has planned for: mechanisms fail too, and the problem is upstream of both documents and mechanisms.
• Whether a run's output is ever merged. Two runs have been abandoned on purpose, so the client has seen nothing render. That decision belongs to #379 on the parent map, and it is not this map's to make — but run 3 produces a third abandoned branch if it stays open.

Out of scope

• Giving DXP-DesignSystem a pipeline. It has no CI of any kind; every green claim about it came from a human running checks locally. Recorded on #373, not this map's job.
• Fixing runs 1 and 2's output. Both are measured artefacts. Five real fixes sit unmerged on 5e683a7 and that is deliberate.
• The run-2 findings that are not instruments — bug(unic-archon-dlc): the re-entry idiom reads before it offers the choice, and Step 8 hardcodes names the Consumer overrides #440, bug(unic-archon-dlc): /specs wrote a PRD requirement the chain's own gates forbid #442 to tech-debt(unic-archon-dlc): the design-contract half is 28% of the /specs command file #451, docs(unic-archon-dlc): remove the five self-describing ADR amendment blocks (Tier A) #453, and since 2026-09-04 #439. They are run 4's change set. Landing them before run 3 would change what the chain produces while it is being measured. bug(unic-archon-dlc): open-pr precedes its gate, and a config-value gate has no safe middle #439 moved here when its grilling turned the fix into behaviour (evidence writes always, a red build opens no PR, /tickets guarantees a parent ticket), all active under gates.build: afk; the maintainer ruled "ok, espera a run 3". Its ADO mirror 43036 stays under Feature 42975 and is no longer part of run 3's count. Its ready-for-agent label (2026-09-04 22:44) means the criteria are settled, not that it may be dispatched: it waits for run 3 to finish AND for bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430 to merge, because its criteria are written against the tree bug(unic-archon-dlc): the config maps commands to nodes and the Boxes name tools, so a check runs by luck or not at all #430 leaves (evidence, /qa's bootstrap). Whoever picks it up after run 3 re-grills the body before dispatching; nothing downstream re-checks this, and it will have sat longer than the few days root AGENTS.md allows a ready ticket. Removed as a child of this map on 2026-09-04 23:05 and unassigned, so the frontier query cannot offer it; it sits with bug(unic-archon-dlc): the re-entry idiom reads before it offers the choice, and Step 8 hardcodes names the Consumer overrides #440, bug(unic-archon-dlc): /specs wrote a PRD requirement the chain's own gates forbid #442–tech-debt(unic-archon-dlc): the design-contract half is 28% of the /specs command file #451 and docs(unic-archon-dlc): remove the five self-describing ADR amendment blocks (Tier A) #453, none of which is on a map, until run 4 is charted.
• /qa's three unrecorded node outputs (e2e.summary, coverage-gate.summary, uat-prep.criteria — emitted, never read, no report node) — filed as #458, not a child of this map. Ruled out of run 3 by the maintainer on 2026-09-03: /qa is not on run 3's path. Found while grilling bug(unic-archon-dlc): a review node produced thirteen findings and recorded them nowhere #438, whose own premise did not survive a direct read — the precheck findings are in report.md § 4 on run 2's branch.