Phase 4: the VRP series — HAR walk-forward, ex-ante vs ex-post, lookahead guard

[NavyBlueCheese]

What was built

• lookahead.py — the guard: any information set containing rows after the decision time raises LookaheadError; null timestamps are refused too. Under test, including deliberate contamination.
• realized/har.py — Corsi (2009) HAR in logs with the two brief-mandated choices: direct 30-day-horizon regression (not iterated one-step) and the lognormal retransformation exp(s^2/2). HAR-RV-CJ via the ABD (2007) continuous/jump split at the BNS 1% level. Walk-forward is expanding-window with a subtlety the tests pin: training rows must have their own forward window closed by the decision date, or the target leaks the future.
• analysis/vrp.py — both VRP definitions, kept in separate columns forever: ex-post (descriptive only) and ex-ante (the tradeable object). The realised leg now includes weekend/holiday gap returns via new first/last log-mid panel columns (conventions rule 5 was silently unmet before this PR — no single intraday window can see the Friday-to-Sunday gap).
• analysis/stylized_facts.py — moments and Newey-West inference; a test demonstrates HAC deflates the t-stat by >2.5x on an artificially overlapped series.

Acceptance tests (brief section 5, Phase 4)

• The lookahead guard raises when fed future data: passing.
• Walk-forward purity: the forecast for date t is bit-identical (rel 1e-12) when all data after t is deleted: passing.
• HAR beats the expanding-mean naive forecast in log-MSE on a persistent synthetic DGP; fit recovers known coefficients; retransformation makes level forecasts mean-unbiased where the naive exp(log-forecast) is >25% biased low.

118 tests, coverage 93.3%, mypy --strict clean.

First real output (auto-extends with the backfill)

With EVZ joined to the currently-ingested ticks, the only fully-covered forward windows are early March 2020 - and the ex-post VRP there is negative (mean -0.0145 annualised variance, positive on 1 of 10 days): implied variance going into the COVID month priced far less than materialised. The canonical crash-month loss, produced by the pipeline unprompted. Ex-ante forecasts appear once ~300 training days of closed windows exist (the backfill is working through 2007 now).