feat(research): measure what a round trip costs on a commission-free … - #697

…venue (#371)

Alpaca Phase C's cost-fidelity half. The PRD makes this a precondition -- "no strategy claim is believed before it" (§6.3, O4) -- and the ordering earned its keep: config.paper-equities.yaml ships taker_pct: 0.0, honestly labelled as awaiting this work, and priced literally that models a FREE venue.

Result: a round trip costs ~2.2bp on Alpaca equities and ~306bp on Coinbase spot, a factor of 141. keel's crypto model is validated in passing (332bp charged vs 306bp measured, conservative by 8%, from an estimator with no knowledge of Coinbase's fee schedule). Its equities pricing is 7.5x too high.

Two findings mattered more than the headline.

THE AGGREGATION CHOICE IS WORTH 22x, AND THE OBVIOUS ONE IS WRONG. Corwin-Schultz two-day estimates go negative constantly. Flooring each pair and averaging the survivors keeps the positive half of symmetric noise and discards the negative half, so the mean converges on E[max(X,0)] > 0 -- it reports a spread that is not there, and a bigger one the more volatile the series is. Worst, therefore, exactly on the asset class this exists to price. The paper's own procedure averages within a month and floors the MONTHLY mean; negatives then cancel inside the block. Same data, same estimator: MSFT 41.8bp -> 1.9bp. The first run of this measurement reported 43.6bp and a 4.1x ratio, and what exposed it was the implausibility of a mega-cap quoting 41.8bp, not a test. Both aggregations stay reachable, because the comparison between them IS the finding.

KEEL READS MSFT AS A THIN ASSET. Cached median daily quote volume is $186M -- about 2% of MSFT's real consolidated volume, which is approximately IEX's market share, and the equities profile runs on the IEX feed. Every ticker sits below the model's $500M anchor, so every one is priced as thin. The candles are fine; the statistic is being asked a question the feed cannot answer. Bounded for cost modelling, unbounded for liquidity screens and admission floors. Filed as #696.

Tests (written first, red before green):

tests/research/test_spread.py -- 28 tests. The load-bearing one is
test_flooring_each_pair_biases_a_quiet_series_up: a seeded random walk
with NO spread at all, where the naive aggregation must read >5x the
blocked one. test_the_overnight_gap_adjustment_is_applied pins §I.B --
equities gap nightly and crypto does not, so an unadjusted estimator would
inflate the equities side of exactly this comparison, in the direction that
flatters keel's existing prior.

Mutation-verified, 12 mutants. One survived -- averaging over positive blocks only rather than all blocks -- because no test had a series with both positive and negative BLOCKS; test_a_floored_block_still_counts_toward_the_average was added and kills it. One further mutant proved equivalent (sum(max(m,0)) and sum(m for m if m>0) over the same denominator) and is recorded as such rather than counted as a survivor.

keel/commands/research.py gains the index row its completeness pin demands.

Refs #371
Refs #696

Claude-Session: https://claude.ai/code/session_01NzuKAe2RVrPt9acVAWjRyL

What & why

Tests-first evidence

• Tests written first, seen failing for the right reason

Gates (all must pass)

• uv run ruff check clean
• uv run mypy clean
• uv run pytest -q green

Scope check

• This PR touches a rail or a default classification — checked means it DOES;
  leave checked only if true, and if so: cite the source and open the discussion
  BEFORE review (CONTRIBUTING.md, "Governance: rulings vs. machinery").
• New dependency added (needs discussion first)