feat(research): the equities DCA benchmark, read per ticker (#371)

[eaitbrahim]

Alpaca Phase C's second half, and the last of #371. It follows the cost-fidelity document in that order deliberately -- the PRD makes that a precondition, and a DCA sleeve priced wrong is exactly the claim it was protecting against.

THE BENCHMARK IS THE MEDIAN TICKER AT +70.62%, NOT THE POOLED SLEEVE AT +145.10%. The run's pre-declared refutation condition -- "a sleeve whose terminal value is dominated by one ticker, making 'beat DCA' a statement about NVDA rather than about equities" -- fired. NVDA returned +449.35% and is 44.8% of the terminal sleeve. Without it the remaining four returned +69.03%, within 1.6pp of the median, which is why the median is the honest summary. Quoting the pooled figure would set a bar one exceptional name built and then credit or blame a strategy for clearing it.

THE COST ARMS BARELY DIFFER, AND THAT IS THE USEFUL PART. keel's 7.5x equities mispricing is worth 0.18 percentage points over five years here; even the fully counterfactual crypto_regime arm (Coinbase's 1.2%/leg on equity bars) costs 3.08pp. A DCA sleeve pays its spread 250 times, one way, on a position it never closes. This is the crypto finding reached from the other side and it is easy to get backwards: cost is levied on the SEARCH, not on the edge. Cheap execution does not make a strategy good; expensive execution makes a mediocre one impossible.

sim/portfolio_sim is deliberately NOT used, and the record says why: it iterates ONE_HOUR bars (_window_bars is history_days * 24, mirroring the live agent's hourly account pass) and the equities profile is ONE_DAY only by configuration. Handing it daily bars labelled as hourly would produce numbers that look right and mean nothing. The accumulation loop is in the driver and calls the SHIPPED Dca rule, filling at the next bar's open plus one-way slippage -- keel's own market-order convention since #258, with the deciding bar always strictly before the filling bar.

Tests (written first, red before green):

tests/test_dca_benchmark_is_read_per_ticker.py -- 6 tests. They pin the PROSE,
because that is where the damage would be done: the headline must name the
median rather than the pooled figure, the concentration must be quantified
BEFORE the first heading, the ex-NVDA row must survive, all three declared
cost arms must still be reported, and the record must refuse to read as a
recommendation. test_one_ticker_really_does_dominate_the_sleeve guards the
premise -- if a re-run ever balances the sleeve, these fail loudly rather
than enforce a caveat that no longer applies.

Mutation-verified, 5 mutants, all killed. One initially survived: gutting the concentration sentence while the table's "44.8%" column kept the pin green. The pin now requires the figure in the opening, since a reader who takes the headline and stops is exactly who the caveat is for.

Closes #371

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)