Declare each rule's parameter space, so the trials budget is derived rather than remembered

[eaitbrahim]

From the 2026-08-23 Jesse-comparison PRD (docs/superpowers/specs/2026-08-23-strategy-api-expressiveness-prd.md §4.1). The one genuinely new finding from that comparison — everything else it surfaced was already filed.

The gap

n_trials — the input to keel's own overfitting correction — is currently a number a human remembers to record.

keel already has the machinery that consumes it:

• research/deflate.py::expected_max_sharpe(n_trials) — the Sharpe expected from the luckiest of N zero-skill trials.
• research/independence.py — "two rules that fire together are one rule counted twice, consuming trials budget twice".
• A hand-maintained ledger (keel/research/ledger.py) recorded through the CLI with explicit DECISIONS/PROVENANCE, plus source comments reasoning per-choice about whether something "increments the trials budget (§73.12)" (analysis/indicators.py:388, analysis/levels.py:114).

So the statistical correction is rigorous and the number feeding it is manual. That asymmetry is the bug.

What Jesse does

A strategy declares its own search space:

def hyperparameters(self):
    return [
        {'name': 'stop_loss_atr_rate', 'type': float, 'min': 0.1, 'max': 2.0, 'default': 2},
        {'name': 'up_length',          'type': int,   'min': 3,   'max': 30,  'default': 21},
        {'name': 'up_coeff',           'type': float, 'min': 0.1, 'max': 3.0, 'default': 0.71},
    ]

Jesse uses it to drive an optimiser. That is not why it is interesting here. It is interesting because the size of the space a sweep could explore becomes a property of the rule rather than of the operator's diligence — which makes the trials count derivable.

Proposal

Add a declared parameter space to Rule — most naturally an extension of the existing describe(), which already returns per-parameter help text, so the names and their documentation live together rather than in two places that can disagree.

Each entry carries at minimum: name, type, and the range considered legitimate for that parameter.

Then:

• The trials arithmetic reads the declaration instead of a hand-entered number.
• A sweep script reads the declaration rather than restating it, so an explored range cannot silently exceed the declared one.
• keel rules describe-style output can show what a parameter is allowed to be, not just what it currently is.

Explicit non-goal — do not build an optimiser

Declaring the space is cheap and improves a correction that already exists. Running an optimiser against a cost-bound engine would manufacture exactly the overfitting the deflated-Sharpe machinery exists to detect — and the 144-cell sweep already demonstrated the shape of that: 8 winning cells out of 864 trials, all in one asset, in a liquidity surge.

The declaration must not become a licence to search it.

Acceptance criteria

• Every registered rule kind declares its parameter space, with ranges that reflect what has actually been considered legitimate.
• n_trials for a rule family is derivable from the rule; no hand-entered count is required for the standard case.
• A sweep reads the declared space rather than restating ranges in a script.
• The declaration is covered by a test asserting it stays in sync with the constructor's real parameters — a declaration that drifts from the signature under-counts trials, which is worse than none.
• The existing ledger's DECISIONS/PROVENANCE recording is unchanged; this feeds it, it does not replace it.
• No optimiser ships as part of this.

Notes

• Related: The 2026-09-30 n=100 pooled review is underpowered: n_eff is about 39, and it can only detect a 20-point edge #427 (the pooled review is underpowered at n_eff ≈ 39) is the other place where a statistical input deserves to be computed rather than assumed.
• The PRD's §6 risk stands: the field is consumed by the correction, and any sweep still records to the ledger with its provenance.