feat: add causal semiconductor regime research observer

[Pigbibi]

Scope

Four new files only: a pure SOXX/SOXL research observation module, synthetic contract tests, and paired English/Chinese documentation. Existing 79 tracked files are byte-for-byte unchanged. No package-level exports, runner/catalog registration, live defaults, dependencies, services, credentials, strategy routing or positions changed.

Base b455e6b / tree 216f02a8247301be1b6bf93965452581ee80a7d4. Exact candidate tree 7ec00258520b3a09a4f573e37045e59e2bb3fa9d.

Behavior

• Separate direction, trendiness and market pressure; quality failures stay unknown rather than becoming risk-off
• Mandatory explicit research configuration; default feature-only output, no calibrated thresholds or probabilities
• Causal as_of/decision_at slicing; hidden future or late revision rows cannot change historical features or input hashes
• Declared session completion/calendar/adjustment/component consistency, conflict and undefined-feature handling, expiry anchored to the source session
• Reuse current canonical JSON and V2 envelope helper; bind repo/entrypoint/config/causal input root without granting trading authority

Verification

• Root independently ran the full candidate suite with network/external-service CLI guards: 365 passed, 13 subtests passed, zero blocked attempts
• Worker: baseline 334 passed, 31 added research tests, 68 focused regressions and 48 full-observation parity cases against reviewed reference
• Ruff 0.16.9, pip check, sdist/wheel and packaged module bytes, exact-base patch/whitespace checks passed
• Existing pandas FutureWarning remains in both baseline and candidate
• The optional AI SDK was installed only in an isolated test environment from the existing exact project pin 65c8bcf432f578f657c12d5dab0e16ebec0bef5a; gateway behavior mocked

Limits

UNVALIDATED_RESEARCH only. qualified_by_declaration checks supplied consistency, not provider truth, real PIT availability, P1 contents, installed/live pins or economic efficacy. SOXX is a signal proxy, not assumed identical to SOXL benchmark/returns. Real publication time proofs, qualified data, OOS/paired shadow and approved consumption remain separate. No automatic activation or consumer repin. Await exact-head CI before normal merge.