feat(scan): add stats.momentum.tsmom@1 intraday momentum detector

[radiusred-cody]

What

New single-leg ANOM scan stats.momentum.tsmom@1 (arity = Single) that measures intraday time-series momentum / return continuation as a tradeable signal — sign + magnitude + significance + directional hit-rate — rather than a bare autocorrelation p-value.

Kernel

Per horizon k, the return series is partitioned into non-overlapping k-blocks (trailing remainder dropped); each block carries its summed return. Consecutive (past, next) block pairs are fit with OLS R_{b+1} = α + β·R_b:

• β = continuation coefficient (>0 momentum, <0 mean reversion; ≈0 under a random walk)
• two-sided Student-t t-stat / p-value (df = m-2)
• directional hit-rate (sign agreement) and the sign(past)·next TSMOM mean

Non-overlapping blocks keep the pairs ~independent so the t-stat is honest (overlapping windows would inflate it via induced autocorrelation). Optional scaling (default on) divides each return by a look-ahead-free trailing volatility before block formation — a global rescale is OLS-scale-invariant, so the normalisation is deliberately time-varying.

Surface

• params: k_values (default [1,5,10,20], each ≥1), scaling (bool, default true)
• effect.metric = "tsmom_continuation", value = β at the selected hold horizon (the k whose positive continuation is most significant; falls back to strongest |t|)
• effect.p_value / effect.effect_size = {hit_rate} at that same k
• effect.extra (parallel per-k arrays): continuation_coefs, hit_rates, k_values, p_values, selected_hold_bars, t_stats, tsmom_means, turnover_per_bar. selected_hold_bars frames which horizon persists and turnover_per_bar = 1/k lets the consumer read natural turnover straight off the finding.
• raw.series = {returns, timestamps_ms}

Registration

Appended alphabetically inside register_anom_scans (Pattern E — registry.rs::bootstrap() untouched); dispatchable on all three surfaces (CLI/MCP/HTTP) via the shared registry. Family mod.rs registration assertion updated.

Tests

• Kernel unit tests (p-value edges, vol-normalize, white-noise ≈0 continuation, AR(1) significant positive continuation, mean-reversion negative, guards)
• Scan unit tests (id/version/arity/schema, envelope shape, AR(1) significance, random-walk near-zero, turnover, cancellation, param validation, hold selection)
• Happy-path integration test + float-free insta schema snapshot (pins catalogue shape — scan-id, metric, effect-size kind, array names + shapes, params — without coupling to exact continuation floats, which the unit tests cover for sign + significance)

rustfmt clean. nextest + clippy validated by CI (the sandbox has no C linker, so the build/test/clippy gates run there).

Notes

• Phase-5 hygiene opt-in (bootstrap / null methods) is intentionally left at the trait default: PhaseScramble preserves the autocorrelation function and so is not a valid null for a linear-continuation detector. Adding this scan to the per-scan hygiene matrix is a separate, centrally-governed decision.

[radiusred-testy]

Changes Requested

CI failed on cargo clippy --workspace --all-targets -- -D warnings — 2 errors, all in the new TSMOM files. Everything else (build, logic, tests, registration, scope) is clean. Fix the two lints below and re-push.

Issues

Error 1 — crates/miner-core/src/scan/anom/tsmom/kernel.rs:130 — clippy::many_single_char_names

error: 5 bindings with single-character names in scope
  --> kernel.rs:130:51

The tsmom_continuation function has five single-char bindings in scope: n (line 131), k (parameter), m (line 150), x (line 166), y (line 167). The existing #[allow(clippy::similar_names)] does not cover this lint. Add a separate allow on the function (or on the module):

#[allow(
    clippy::many_single_char_names,
    reason = "n/k/m are canonical OLS/TSMOM names; x/y are the standard regressor/regressand pair"
)]
pub(crate) fn tsmom_continuation(returns: &[f64], k: usize) -> Result<TsmomResult, String> {

Error 2 — crates/miner-core/src/scan/anom/tsmom/mod.rs:188 — clippy::similar_names

error: binding's name is too similar to existing binding
  --> mod.rs:188:17
note: existing binding defined here
  --> mod.rs:129:9 (req: &ScanRequest)

res (line 188 in the k-loop) is flagged as too similar to req (the run parameter, line 129). Either rename to something unambiguous (k_result, tsmom_k) or add #[allow(clippy::similar_names)] to the run method. The #[allow(clippy::too_many_lines)] already on run does not cover this lint — it needs its own entry:

#[allow(
    clippy::too_many_lines,
    reason = "…",
)]
#[allow(
    clippy::similar_names,
    reason = "res / req are standard kernel-result / scan-request shorthands"
)]
fn run(…) { … }

or combined:

#[allow(clippy::too_many_lines, reason = "…")]
#[allow(clippy::similar_names, reason = "res is kernel result; req is the scan request — distinct roles")]
fn run(…) { … }

Checks already passed

• cargo build --workspace ✓
• Scope, logic, tests, registration (Pattern E, alphabetical), conventional commit — all pass
• No security issues, no private research data
• Will approve immediately once the two lints are suppressed and CI goes green

[radiusred-testy]

Approved

Re-review complete. Both lint issues from my earlier CHANGES_REQUESTED are resolved:

• tsmom_continuation carries #[allow(clippy::many_single_char_names, reason = "...")] directly on the function with a clear justification — ✅
• k-loop variable renamed res → k_result throughout the dispatch path — ✅

Checks run

• Scope vs RAD-3839: new stats.momentum.tsmom@1 scan only — no scope creep ✅
• Conventional commits: feat(scan): (new user-visible feature) + style(scan): (lint fix) — both correct ✅
• CI: green on fix commit 76716a3d (12m47s push + 12m6s PR runs) ✅
• Private research data: none ✅
• Security: no exploit paths, no credential exposure ✅
• Lint suppression: scoped with reason = attributes, not blanket silencing ✅

PR is clear to merge.