A provider 429 or 529 is a transient account condition, not a session
defect. The recovery loop retried it immediately, so every resume burned
in milliseconds while the account stayed throttled, and the campaign's
declared retry then repeated the same futile sequence. Seven executions
of the first qualified 3x3 cohort died this way.

Throttle interruptions now wait on an escalating schedule under their
own time budget (STACK_BENCH_PROVIDER_THROTTLE_MAX_WAIT_MINUTES, default
300) instead of consuming the bounded interruption retries, resume the
exact paid session when one exists, and restart from the existing files
when the first request itself was rejected. Waits are logged, recorded
in the session interruptions, and reported in run setup so duration
comparisons cannot blame a backend for the account's usage window.

The adapter deadline also rises above one full coding session plus the
throttle budget: the old 75-minute value sat below the 123-minute
session bound, so the supervisor's SIGTERM pended through the
synchronous container call and reported live sessions as failures.
Adapter contract 1.11.0 -> 1.12.0.

Campaign run validation canonicalized selection fields straight from the
artifact, so a run that omitted one crashed with 'definition plan at $
is not canonical JSON data' instead of naming the diverging field. A
missing field now registers as an ordinary identity mismatch.

Also repairs the test debt left by the modular example migration: the
level fixtures now carry the exact planned selection a valid run.json
records, and the condition-version expectation matches the example.

Four operations the grader calls by fixed name were never stated to the
builder: updateCartQuantity (PATCH /api/cart/:itemId) at L1, and
adminTransferStock, adminChangePrice, and cancelOrder with their routes
at L2. Apps passed those checks only when they guessed the reference's
name — a lottery REST convention wins more often than reducer naming, so
it biased L2 against SpacetimeDB. In the running 3x3 cohort, spacetime
attempts enforce authorization correctly (530 refusals where the reducer
exists) but lose 201a/201b to a 404 on a name they were never told.

New draft releases state the names the way L1 already states its six:
the cart feature text names updateCartQuantity, and the three L2 packs
gain pack-owned testing-call tables. Recipe diff proves scoring,
fixtures, and execution are byte-identical; only task fragments were
added, and the L2 calibration is correctly invalidated pending a fresh
qualification of l1-modular@2.5.0 and l2-standard@1.6.0.

A 404 on a named action now also names the required operation in the
failure text, so a repair round can create the missing endpoint instead
of chasing authorization. Verdict logic is unchanged.

The reseed/restart probe reads GET /api/items and requires a JSON object
with an items array, but no prompt stated the endpoint or its shape - a
server-based app returning a bare array or routing its catalogue
elsewhere fails as an application startup-seeding failure it was never
told how to avoid. Same undeclared-surface class as the named testing
operations; found by completing that audit. The draft l1-modular@2.5.0
now states the endpoint as recipe framing; SpacetimeDB apps are exempt
because their probe goes through the module.

A postgres attempt seeded only on first install, came back empty after
the harness's database reset, and scored 0/58 without grading. The task
now states the behavior the reseed probe verifies: the server seeds its
starting data at startup whenever it is missing. Draft 2.5.0 fragment
only; qualified releases untouched.

The landing table led with cost to correct - the secondary metric - and
rendered a completed campaign's status as an empty cell while stale
history shouted for attention. Each stack cell now leads with the
first-build score, shows the repaired score and median spend beside it,
and stars a cost the campaign's own detail view would refuse to compare.
Completed says so. History that needed attention days ago archives out
of the default view behind a toggle, campaign keys distinguish
identically titled rows, and running campaigns pin to the top with a
per-attempt strip: phase, latest score, elapsed time, spend so far.

Refresh now reconciles keyed rows in place instead of replacing the
table, so an element under the pointer survives the poll; the whole row
is the click target. Campaign detail is URL-addressable
(#campaign=<key>), opens facts an operator previously read plan.json
for (agent, model, exact recipes, pinned image digests), and each
collapsed attempt answers first -> final, rounds, duration, and spend,
with the still-failing checks as chips.

The model distinguishes a first grade that aborted before scoring
(seed probe, startup failure) from a scored zero - an app the grader
never scored did not score 0 percent - and level failure lists now use
the authoritative final appFailures instead of falling back to
first-build misses on repaired attempts. Levels gain durations; plans
gain identity facts.

Medians answer which stack won; an operator watching a cohort needs the
nine runs. A campaign row now expands in place into one row per attempt
- stack and repetition, state or live phase, level reached, first and
latest score, rounds, elapsed time, spend - updating on every poll.
Running campaigns start expanded, the choice survives refreshes, and
the title button still opens the evidence modal. Replaces the
running-only chip strip, which compressed the same facts into prose.

The page led with a history table; the operator's questions - is it
healthy, how far along, what just happened, is anything stuck - had no
surface. Three zones now answer them in order.

Now: each running campaign is a card with attempts done, burn against
the declared ceiling, and elapsed time; each attempt is a lane with a
score bar, live phase, a score-per-round sparkline, and an event line.
Two failure smells the operator previously diagnosed from logs are
automatic: a trailing run of identical grades flags 'flat N rounds',
and a running attempt whose output file has gone quiet flags
'no output Nm'. Scores flash on change and a liveness tick shows the
last refresh while anything runs.

Latest verdict: the most recent completed campaign renders as a
dot-and-whisker row per stack on a shared first-build axis, repaired
score and median spend beside it, with the aborted-grade and exclusion
footnotes attached to the claim they qualify.

History: the table, demoted below both, without running campaigns.

Model: parseRunProgress keeps the whole grade series instead of only
first and latest, and attempts expose when their log last changed.

The history table forced three stack columns onto every campaign, so a
single-stack smoke run rendered as a comparison with two holes and a
starred median. History's job is identification: each row now carries
the participating stacks beside the key, one adaptive headline showing
first->repaired only for stacks that actually graded, and the
campaign's total burn - always well-defined - instead of per-stack
medians whose comparability belongs to the verdict and detail views.

The detail view stops being a floating window: it is a full-height
right side panel that closes on outside click or Escape, keeping its
deep link. A top load bar reflects requests actually in flight.

The adaptive headline shipped without its layout, so per-stack figures
fused into unreadable run-on text. Each history row's result is now a
two-column block - stack labels, aligned first->repaired figures - with
the campaign's total burn on its own line.

Three redesigns tried to make a summary stat legible inside a list
cell; the right answer was that it does not belong there. A history row
now identifies the campaign - title, key, participating stacks, status,
recency - and nothing else. The numbers live where they can be read
honestly: the verdict card, the row's per-attempt expansion, and the
detail panel.