feat(market-data): UK (LSE .L) equity market support

[cgycorey]

Review fix: yfinance currency (91aa401)

Finding: the yfinance fallback inferred currency from the .L suffix and ÷100'd every LSE line — but .L is not uniformly GBp. Verified against Yahoo get_history_metadata()['currency'] (2026-08-29): VOD.L=GBp, HSBA.L=GBp, IUSA.L=GBp, SGLN.L=GBp, VUSA.L=GBP, VUSD.L=USD. A backtest via the yfinance path would corrupt every absolute quantity (initial-capital sizing, lot rounding, min-notional, cross-market composite weights) while returns survive.

Fix: the yfinance path now reads Yahoo's declared currency (Ticker.history_metadata['currency'], same field the direct-Yahoo path already passes to scale_pence_to_currency) and scales only on GBp/p. A missing currency fails closed — "do not scale", never "assume pence" from the suffix. GBp names (VOD.L) still ÷100.

Tests (in test_yfinance_uk_scale.py): GBP-quoted .L (VUSA.L) unscaled; USD-quoted .L (VUSD.L) unscaled; absent currency unscaled (fail-closed); GBp still scaled.

Branch rebased onto origin/main (e90b6c6c); 16+1 commits.

References

• Issue: [Feature] UK equities (LSE .L) market support in market data + backtest routing #1205 — [Feature] UK equities (LSE .L) market support
• Tax rule sources: gov.uk "Tax on buying shares" + HMRC Stamp Taxes manual STSM014030 (FA86/S99(13) rounding), STSM041270 (AIM exemption), STSM101060 (ETF exemption)

Scope

UK LSE (.L) equity market support per #1205. Full parity: routing, classification, settlement currency, price units, engine dispatch, statutory SDRT, plus the journal/shadow and financial-statements downstream consumers. (Irish ISE .IL explicitly out of scope.)

What this adds

UK equities (LSE .L) currently fall through _SOURCE_PATTERNS to the tushare default, get classified a_share, and die on the China fallback chain with a misleading _unresolved (several seconds of network attempts, indistinguishable from a dead ticker). This makes UK a first-class market with full parity to Canada/US:

1. Routing — _SOURCE_PATTERNS: .L → yahoo; FALLBACK_CHAINS: uk_equity: [yahoo, yfinance, local]
2. Classification — _MARKET_PATTERNS → uk_equity; _detect_submarket → uk
3. Settlement — code_currency → GBP (was UNKNOWN:uk_equity)
4. Engine — composite/runner dispatch → GlobalEquityEngine(market="uk"); never silent CryptoEngine fallthrough (2x set literals in runner, + _MARKET_TO_SOURCE)
5. SDRT cost model — 0.5% Stamp Duty Reserve Tax on purchases, penny rounding with exact-1/2p rounds-up (FA86/S99(13)), Decimal ROUND_HALF_UP; config key uk_stamp_tax; trade-side detection is is_open-aware (opens: direction>0=buy; closes: -direction — closing a long is a sale, covering a short is a purchase, matching base.py's position-side close path); validated against HMRC manual worked examples (13.4547->13.45, 13.455->13.46)
6. Volume units — yahoo/yfinance loaders declare uk_equity: shares (was null/undeclared, breaking the fix(data): declare per-market volume units in market-data provenance #1065/[Bug] A-share volume unit inconsistency across fallback chain (lots vs shares, 100x) and missing unit metadata #1062 provenance contract)
7. Price units — Yahoo serves LSE names in pence (GBp): VOD.L was ~117p, HSBA.L ~1500p; the engine priced a GBP 1.17 share at GBP 117 (100x notional error on positions and SDRT). Normalize OHLC /100 at the loader boundary (chart-meta currency == "GBp" on the Yahoo path, suffix rule on the yfinance fallback) so every consumer sees GBP values matching code_currency; provenance declares currency_conversion: "GBp->GBP (/100)"; loader cache version bumped v4->v5 (pre-normalization parquet never served again, [Bug] A-share volume unit inconsistency across fallback chain (lots vs shares, 100x) and missing unit metadata #1062 precedent)
8. Journal/shadow consumers — _infer_market_from_symbol: .L → "uk" (was "other", silently degrading shadow-analysis price features to NaN); shadow _MARKET_KEY_MAP/labels map uk → uk_equity, "UK equity"
9. Financial statements — `.L` routed to Yahoo quoteSummary in `get_financial_statements` (income/balance/cashflow annual histories + financialData/defaultKeyStatistics indicators, `{raw,...}` flattened to numbers); was rejected with a suffix error. `get_fundamentals` stays US-only by contract (SEC XBRL PIT panels — Yahoo has no filed-date PIT history), failing loud for UK

Verified

• Live: VOD.L close now GBP 1.1715 (was 117.15), HSBA.L GBP 15.184 (was 1514.6) — yahoo direct
• Economics: 1,000 sh VOD.L -> GBP 1,171.5 notional, SDRT GBP 5.86 (was 585.75) — the 100x unit bug is gone end to end
• 386 affected tests pass locally (UK suites: yahoo client/loader, yfinance scale, market-data provenance, support + backtest smoke, registry, journal, shadow, financial statements, stock profile; plus upstream settings/source-order suites); CI green on all 9 checks (Python 3.11 + 3.14 tests, package, hash locks, Windows lifecycles)

Known scope notes

• Fee model: SDRT is the only exchange-mandated UK cost modeled; broker commissions are caller-config (same as US zero-commission default)
• SDRT exemptions (UCITS ETFs, eligible AIM shares, gilts, new issues) are the caller's concern, not per-symbol detection
• Scope note: Yahoo's Irish ISE surface is .IR/EUR (BIRG.IR serves EUR data); routing it requires a separate market class (EUR currency, unverified Irish tax rules) — intentionally out of scope. This PR covers LSE .L only

Follow-ups (not included)

• Live market-open specs (MARKET_SPECS has no London session; missing market fails loud, never silent)
• Symbol-search UK suffix hints

[warren618]

Thanks — the UK work itself reads well, and the parts I checked (SDRT on the purchase side, GBp→GBP normalisation at the loader boundary, .L routing through the fallback chain, statements / profile / journal) are the right shape.

I can't merge it as it stands, because it isn't only UK support. The branch is up to date with main at 80ffdda4, which already carries #1232 and #1233 — and it still changes src/live/runtime/ by +502/−54, a second implementation of the halt-sweep and flatten-latch work that landed on 2026-08-27:

• _halt_episode goes from "the per-broker sentinel wins" to "the newest sentinel wins", and the latch record accumulates an episodes list instead of binding one.
• mark_sweep_fired moves from before the sweep (runner.py:585 on main) to after it, gated on a new side_effects_attempted flag.

The second one changes the kill switch's no-retry rule. It may well be the right change, but it cannot be reviewed inside a market-data PR, and the diff that would carry it is titled "UK equity market support".

The underlying point is legitimate and I want the fix. I reproduced both gaps against current main:

1. A broker read that returns an error envelope ({"status": "error", ...}) is iterated as a mapping, so the sweep dies with AttributeError: 'str' object has no attribute 'get'. fix(live): fail closed on broker error envelopes during the halt sweep #1232 taught the submit responses about envelopes; the reads never learned.
2. Because mark_sweep_fired runs before flatten_and_cancel, a sweep whose reads both fail still latches — so the kill switch's action is silently skipped for that episode, and neither a later tick nor a restart retries it.

So: please rebase onto main and drop 10a438e6, bd20789c, 0ea29b36 and bd93142 from this branch, leaving the UK commits. Send those four as their own PR against main, stating those two behaviours as the thing being fixed with a test for each, and I'll review it promptly. Splitting also means the UK support isn't held behind a live-safety review.

[warren618]

Thank you for doing the split — the four live commits are gone, the scope is LSE-only, and CI is 9/9. I also want to note the _LOADER_CACHE_VERSION bump to v5: pre-normalization pence entries becoming unreachable is exactly right and easy to forget.

I cannot merge it yet, because the yfinance path infers the currency from the ticker suffix:

if is_gbp_pence_symbol(symbol):          # ^[A-Z0-9&.\-]+\.L$
    normalized, _ = scale_pence_to_currency(normalized, "GBp")

.L is not uniformly GBp. Live Yahoo get_history_metadata()['currency'], checked 2026-08-29:

ticker	currency
VOD.L	GBp
HSBA.L	GBp
IUSA.L	GBp
SGLN.L	GBp
VUSA.L	GBP
VUSD.L	USD

So a backtest on VUSA.L or VUSD.L through the yfinance fallback gets every OHLC price divided by 100. Returns survive, since the whole book is scaled together, but every absolute quantity does not — initial-capital sizing, lot rounding, min-notional checks, and any cross-market composite run that weights this sleeve against a correctly-scaled one.

The direct-Yahoo path in this same PR already does the right thing: it passes the chart meta's real currency into scale_pence_to_currency, so a GBP or USD line falls through untouched. yfinance exposes the same field. Please read the declared currency there too, and fail closed when it is absent rather than falling back to the suffix — a missing currency should mean "do not scale", not "assume pence".

Worth a test with a .L symbol whose source currency is GBP and one whose currency is USD, both asserting the prices come through unscaled. Everything else in the PR I am happy with.