rsi_meanrev's gross edge does not survive being made to fire: the last live lead closes negatively

[eaitbrahim]

Follow-up to #251 / #252, answering the single open question that document left.

What #252 left open

rsi_meanrev had the best gross-edge distribution of the three shipped signal rules — median
gross PF 1.1631, 58% of assets gross-positive — while reaching the promotion floor on zero
of 24 assets (median n=38 against min_trades=100). #252 called it "the one live lead" and framed
the question as:

Does its gross edge survive being made to fire more often? If selectivity is what creates the
edge, the rule is unpromotable by construction. If the edge survives to n≥100, it is the only
route by which any rule the codebase ships reaches its own promotion floor honestly.

What was run

A monotonicity test, pre-registered in the script itself this time — the defect §7 of #252
recorded was that its declaration lived in a dispatch brief rather than in the file.

• Variable axis: oversold ∈ {20, 25, 30, 35, 40}, the only measured frequency lever (×2.18 and
  ×3.93 per step, against ×1.186 for support_proximity_pct).
• Held: overbought=80 (an exit control — moving it changes outcomes, not just counts, which
  would make the slope uninterpretable), support_proximity_pct=0.005, all else default.
• oversold=20 reused from docs(experiments): three rules, 24 assets, zero free parameters — the viable intersection is empty #252's rows as the anchor → 4 new levels × 24 assets = 96 combos.
• Conditional proximity arm declared in advance with its trigger and reporting rule.

Result

Conditional arm did not fire — 21 of 24 assets reach n≥100 at oversold=40.

	median gross PF	gross>1
oversold=20 (median n=38)	1.1631	14/24
every cell with n≥100	0.8938	11/76

Net > 1.0 at n≥100: 0 of 76 cells, at any fee including maker.

The edge evaporates exactly when the rule becomes measurable. The 1.1631 that made this rule look
like the best of the three is what n=38 looks like.

Method notes against ourselves

1. The pre-registered statistic was underpowered for its own question. A linear slope was
   declared; the phenomenon is a threshold effect. Mean slope −0.0386/100 trades, 15/24 negative —
   directionally right but noisy, with most of its 0.35 stdev coming from two assets that barely
   trade (PAXG-USD n=12→145, WLD n=1→71). The partition on the pre-declared n≥100 floor is far
   cleaner and is reported alongside it.
2. The design's monotonicity assumption is violated on 3 of 24 assets. UNI collapses 309→9
   between oversold 30 and 35; BTC 181→80. Under investigation — likely an unclosed position
   blocking subsequent entries.

Deliverables

• docs/experiments/2026-08-12-rsi-meanrev-scale-vs-selectivity.md
• The harness, with its pre-registration in the docstring
• Trials-ledger row