fix(maf): en vakt som gikk inert i STILLHET, funnet ved aa loefte pinnen (F15, ORDRE 20260829T155150Z)

MAF core 1.9.0 -> 1.16.0, orchestrations 1.0.1 -> 1.1.1. De to kan ikke loeftes
hver for seg: orchestrations 1.1.1 krever selv core>=1.15.0.

Iron Law: vakt-testen kjoert ROED mot 1.9.0 (2 failed) FOER pinnen ble roert.
Gulvet bor i EN konstant og pyproject-asserten deriverer sin streng fra den.

NEVNER: 16 private/ugaranterte former, derivert fra repoets EGNE siteringer,
alle 16 sjekket mot begge versjoner, 2 endret seg. Kjent-positiv: MiddlewareFailure
flippet NO -> YES. KP-kandidaten _compaction.py ble FORKASTET (teller 0 i begge,
diskriminerer ingenting).

DEN FARLIGE ENDRINGEN er den ordren navnga - formen som fortsatt importerer, men
har flyttet semantikk i stillhet. En park skriver naa TO checkpoints og bare EN
baerer plan-review-typen, saa en feildeklarert _ALLOWED_CHECKPOINT_TYPES toemmer
ikke lenger listingen: den taper nOEyaktig den checkpointen som betyr noe,
get_latest returnerer den ANDRE, og _parks `latest is None`-vakt passerte mens
kjOEringen svarte rc=0 og skrev et spOErsmaal som aldri kan baere svaret. Vakten
sjekker naa EGENSKAPEN den alltid mente (request_id in pending_request_info_events
- et DEKLARERT felt) i stedet for symptomet som pleide aa innebaere den, og fjerner
dermed en privat avhengighet i stedet for aa legge til en.

ExperimentalWarning-paret P4 pkt. 2 betalte for aa BEHOLDE er borte fordi MAF
sluttet aa sende det: _feature_stage.py emitterer ved FOERSTE BRUK, ikke ved import.
Goldenens stderr regenerert som BESLUTNING (fire -> to linjer); site-packages-
maskeringen BEHOLDT (spannet er ubebodd, ikke pensjonert).

Load-bearing MAALT mot HELE suiten, gronn kontroll 1089/5, stdout BYTE-UENDRET
(ea8c534773acdbe41ae68f2c55724d69aaf8be4f): M1 revert av vakten -> 1 rod.
EN mutasjon ble IKKE rod og staar som aerlighets-grense, ikke som gate: spikens
checkpoint_ids[-1] er rekkefolge-avhengig (Path.glob), altsaa flaky.

Rapport: docs/2026-09-02-f15-maf-pinnen.md

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/portfolio_optimiser/explore.py

@@ -1534,18 +1534,27 @@ async def _park(
     ``get_latest`` rather than the last entry of a listing: it picks by timestamp
     (``_checkpoint.py:424``), while the listing's order is whatever ``Path.glob`` returned.
 
-    An EMPTY listing is refused (``CheckpointUnreadable``) rather than parked around. That is the
-    one place this layer is louder than the framework: a blocked deserialisation is swallowed into
-    a warning upstream, so the alternative to raising here is a question file whose answer can
-    never be applied — an exploration that fails as an absence, days later, to somebody who has
-    already written their answer.
+    The checkpoint named here must be the one that carries THIS question, and that is asserted
+    against the declared ``pending_request_info_events`` field rather than inferred from the
+    listing being non-empty. **The weaker form went inert under core 1.16.0 (F15, measured):** a
+    park now writes TWO checkpoints, and only one of them carries a plan-review type. A
+    mis-declared ``_ALLOWED_CHECKPOINT_TYPES`` therefore no longer empties the listing — it drops
+    exactly the checkpoint that matters, ``get_latest`` returns the OTHER one, and the old
+    ``latest is None`` guard passed while the question file named a checkpoint that can never bear
+    the answer. That is the silent dead question this layer exists to refuse, so the guard now
+    checks the property it always meant rather than the symptom that used to imply it.
+
+    Refusing is where this layer is louder than the framework: a blocked deserialisation is
+    swallowed into a warning upstream, so the alternative to raising here is a question whose
+    answer can never be applied — an exploration that fails as an absence, days later, to somebody
+    who has already written their answer.
     """
     latest = await checkpoint_storage(checkpoint_dir).get_latest(workflow_name=workflow.name)
-    if latest is None:
+    if latest is None or str(request.request_id) not in latest.pending_request_info_events:
         raise CheckpointUnreadable(
-            f"the exploration reached a plan review but no checkpoint could be read back from "
-            f"{checkpoint_dir!r}: without one the review can never be resumed, so it is refused "
-            f"here rather than written as a question nobody can answer"
+            f"the exploration reached a plan review but no checkpoint carrying it could be read "
+            f"back from {checkpoint_dir!r}: without one the review can never be resumed, so it is "
+            f"refused here rather than written as a question nobody can answer"
         )
     review = request.data
     return ParkedExploration(

src/portfolio_optimiser/tracing.py

@@ -218,7 +218,7 @@ def tracing_notice(setup: TracingSetup) -> str | None:
 
     ``None`` when tracing is off — omission, never an empty row (``mandate.announce``'s rule, the
     one ``cost_baseline_notice`` and ``skipped_links_notice`` follow). Here it is load-bearing past
-    style: the pinned demo stderr is four lines, and a "tracing: off" row would have made it five.
+    style: the pinned demo stderr is two lines, and a "tracing: off" row would have made it three.
 
     The OTLP form names the VARIABLE beside the value, because the variable is what the operator
     edits — and prints one row per declared endpoint rather than only the first, so a run exporting