fix(sleep): keep the gate non-empty when every task hashes into test

[hylin (linhongyu510)]

Fixes #271.

The gap

assign_splits in skillopt_sleep/mine.py says right in its own comment
"Guarantee val (the gate) is non-empty when we have >=2 real tasks", but
_promote_one only pulls from train (to top up val) or from val (to top
up train), never from test. If every real task's hash bucket lands in
[val_cut, test_cut), both train and val start empty, so both guarantee
calls have nothing to promote and silently no-op.

run_sleep_cycle then finishes with gate_action='reject', edits=0, no
error, no warning, and holdout_leaked does not flag it either. The night
looks like it ran and learned nothing, with no signal saying why.

As #271 notes, this is not contrived: for a small nightly batch (2-5 tasks is
realistic for a solo user), any test_fraction above roughly 0.5 makes it a
matter of when, not if.

Reproduction on 79124b37

from collections import Counter
from skillopt_sleep.mine import assign_splits
from skillopt_sleep.types import TaskRecord

tasks = [
    TaskRecord(id=f"task-{i}", project="/p", intent="x",
               reference_kind="exact", reference=f"a{i}", split="", origin="real")
    for i in range(5)
]
assign_splits(tasks, val_fraction=0.10, test_fraction=0.80, seed=42)
print(Counter(t.split for t in tasks))

Before: Counter({'test': 5}) — val and train both empty, no warning.
After: Counter({'test': 3, 'val': 1, 'train': 1}) plus a WARNING naming the
fractions and seed.

The change

Two small, additive adjustments in assign_splits:

1. _promote_one now returns whether it promoted anything, so a no-op is
   distinguishable from a promotion.
2. Each guarantee keeps its existing preferred source and only reaches into
   test when that source is empty. #235's stability guarantee is untouched
   in the normal case — hash-assigned test tasks are reassigned only in the
   degenerate case where the alternative is a dead cycle.

Topping up train prefers test over a single-task val, because taking
the only val row would re-empty the gate that was just filled. When there is no
test slice at all, it still falls back to val exactly as before.

Borrowing does spend a held-out task, so it is logged via
logging.getLogger("skillopt_sleep").warning(...) — the module's existing
convention — naming the task count, both fractions and the seed, and pointing
at test_fraction as the knob to lower.

This is the shape #271 proposed. I also hit one case the issue does not
mention: when dream tasks are present they occupy train, so the second
guarantee is already satisfied and only the val guarantee fires. Filling val
from test still has to happen there, and val stays real-only.

Tests

Added to Pass1ApproachBAssignSplitsInvariants, next to the existing #235
stability test:

• test_val_guaranteed_when_every_real_task_hashes_into_test — subtests over
  2..10 real tasks
• test_degenerate_split_warns_that_test_was_spent
• test_dream_train_does_not_mask_an_empty_gate — val stays real-only; dream
  stays train
• test_normal_split_does_not_borrow_from_test — asserts no logger is even
  constructed on the healthy path
• test_single_real_task_is_left_alone — the >=2 guard still applies

Reverting mine.py alone turns 7 of these red, so they pin the behavior rather
than restate it.

Verification

• python -m pytest tests/ — 1501 passed, 12 skipped, 362 subtests passed
• python -m ruff check skillopt_sleep/mine.py — clean

tests/test_split_hardening_2x3.py has one pre-existing ruff I001 import-order
finding on main; I left it alone to keep this diff minimal.

[hylin (linhongyu510)]

Closing as a duplicate — I missed #272, which was opened three days earlier for the same issue, touches the same two files, and lands the same fallback (prefer the non-test pool, reach into test only when both required pools cannot otherwise be filled, and warn when test coverage is spent). Apologies for the noise; #272 should be the one reviewed.

One detail from my run that may be worth folding into #272 if it is not already covered: when dream tasks are present they occupy train, so the train guarantee is satisfied and only the val guarantee fires. #272 gates the val top-up on train_count >= 2 counting all tasks including dream, so a batch with dream rows plus an all-test real hash should be worth an explicit assertion that val is still filled from a real task and never from a dream row.

[hylin (linhongyu510)]

Duplicate of #272.