fix(py): restore 68 batch indicators silently degraded to streaming classes (84cu)

[lavs9]

Summary

qw.supertrend was a class. qw.rsi was a function. Same registry, same generator, no error, no warning — and the difference was pure luck.

The generated TA registry added in 0.7.0 (5ipk.3) derives native batch symbol names with pascal_to_snake() (SuperTrend → super_trend), but the export_*! macros emit pub fn [<$name:lower>] (SuperTrend → supertrend). Every multi-word name missed. The 44 single-word names (rsi, sma, atr) survived only because pascal_to_snake("Rsi") == "rsi".

_resolve_ta_binding treated the miss as a fallback and returned native_streaming — a class with a completely different calling convention. 68 of 221 indicators were affected. 0.6 had no registry layer and bound native symbols directly, which is why it was unaffected: this is a 0.7 regression.

Why this matters beyond the bug

This is the GMM::fit() disease from planning/BRUTAL_REVIEW_2026-07-07.md: a public API reporting success while doing something other than what the caller asked. The cost here was not just wrong behavior — the silence manufactured a false explanation. An agent investigating qw.supertrend being a class concluded 0.7 had intentionally moved SuperTrend to the .ta expression namespace, and planned a downstream compiler refactor around a two-line codegen bug.

The test could not catch it by construction:

assert (entry.get("native_batch") or entry.get("native_streaming")
        or entry.get("polars_method")), slug

It asserted a name was present, never that it resolved. native_batch: "super_trend" passed cleanly.

Changes

• scripts/generate_api_stubs.py — derive batch names with .lower(), matching the macro. Repairs 61.
• scripts/api_slug_aliases.json — 4 stale snake_cased names (fm_demodulator, fourier_series_model, my_rsi, precision_trend_analysis); 3 aliases declaring batch exports that do not exist now fall through to their polars methods (linreg, oc2, true_range); sr_monitor dropped SrInteractionMonitor, a class that exists only as a string in metadata_registry.rs and was never exported to Python.
• quantwave-py/python/quantwave/__init__.py — _resolve_ta_binding raises ImportError when a declared native_batch symbol is absent from the build, instead of silently substituting the streaming class. A native_batch of None still falls through to streaming/polars unchanged (intentionally streaming-only indicators keep working).
• tests/python/test_api_surface.py — assert every declared native symbol resolves against the compiled module, plus a regression test that multi-word slugs bind as batch functions. This test found a 69th mismatch (sr_monitor) during development.
• _ta_registry_generated.py — regenerated: 68 insertions, 68 deletions, exactly the predicted blast radius.

Compatibility

None required. qw.supertrend(period=10, multiplier=3.0, high=..., low=..., close=...) again returns list[SuperTrendResult] — the exact 0.6 calling convention.

Verification

Run against a real 0.7.0 environment:

• 221 slugs · 0 declared-batch-bound-as-class · 0 registry names absent from build
• 127 bound as batch functions; 2 legitimately streaming-only
• tests/python/: 179 passed. The 2 failures (test_sma_parity, test_correctness_precheck_runs) are pre-existing and reproduce on a clean tree via git stash.

Follow-up (not in this PR)

test_sma_parity fails on df.lazy().ta.sma(...) with 'LazyFrame' object has no attribute 'ta'. The .ta namespace is registered on Expr, not LazyFrame — the test encodes the same wrong mental model that misled the agent investigating this bug. Worth fixing separately.

Closes quantwave-84cu

🤖 Generated with Claude Code