fix(keys): the per-arm predictions artifact leaves the trader's prefix (alpha-engine-config-I9822)

[cipher813]

Closes the open half of alpha-engine-config-I9822.

What was wrong

predictions/ held two artifact shapes:

Shape	What it is
predictions/{trading_day}.json	the trader's champion serving feed — the contract named in crucible/slots/__init__.py and in this repo's AGENTS.md
predictions/{arm~seg}/{trading_day}.json	the per-arm artifact

A consumer doing store.list_keys("predictions/") saw both and had to discriminate by path depth. Nothing did — which is what kept this a design risk rather than a live defect, and also why it would have been found the hard way.

What changed

• The per-arm shape moves to arm_predictions/, behind a named ARM_PREDICTIONS_PREFIX.
• predictions_key(trading_day) is declared, so the trader contract has one source for its key instead of living only in prose. The key builder only — writing that artifact is the M slot's job and stays on -I10129. It is declared here because arm_predictions_key cannot be tested against a collision with a key that does not exist.

Moved while the prefix was EMPTY, and that is the point

Measured 2026-09-11: zero objects under predictions/ in the production store.

aws s3 ls --recursive s3://alpha-engine-crucible-v2/crucible/predictions/   ->  0

So this rename orphans nothing. The same change made after either shape had been written would have stranded those objects at an address no code resolves — which is -I10498, filed the same day for a feature layer stranded exactly that way by a content-hash move (532 sessions at vf795db2b5049, unreachable, while live code resolves v6df3c0a27b70 holding 7). The cheapest moment to separate two artifact shapes is before either exists.

Deliverable 1 was already done

The issue's first deliverable — move arm_predictions_key out of slots/inputs.py into keys.py — is already true on main: it is at crucible/keys.py:317 and slots/inputs.py:102 imports it from there. Only deliverable 2 remained. Noted so the issue's closing comment does not claim work that a previous change did.

The guard is proven red, not assumed

Reverting ARM_PREDICTIONS_PREFIX to "predictions/" and re-running:

FAILED TestPredictionsPrefixesCannotCollide::test_the_two_shapes_live_under_different_prefixes
FAILED TestPredictionsPrefixesCannotCollide::test_a_listing_of_the_feed_prefix_cannot_return_an_arm_artifact

Four assertions in total, including the adversarial case the shared prefix allowed — an arm whose id looks like a trading day — and one asserting arm_key_segment stays invertible and still refuses an arm_id that already contains ~, which is the issue's stated gotcha.

Two things found on the way

1. crucible.keys.__all__ was out of alphabetical order before this change. The repo's own TestAllIsSortedAndDeduplicated caught my addition and the pre-existing disorder together. Re-sorted and de-duplicated: 79 entries.
2. The tests went into tests/test_keys_prefixes.py, the file that already owns this class of assertion, rather than the tests/test_keys.py the issue named — which does not exist. Creating it would have made a second home for one question. Reality over the issue's guess.

Test plan

Full suite green before opening, under the repo's own invocation:

uv run --frozen pytest -q --ignore=tests/acceptance   ->  all passed

Worth recording: a bare python3 -m pytest in a worktree reports 8 spurious failures in tests/test_deploy.py with PackageNotFoundError: No package metadata was found for crucible — the laptop's global interpreter, not the repo environment. uv run --frozen is the invocation that measures the repo; the bare one measures the laptop.

Prepared by: Claude Opus 5 (1M context) via Claude Code