C4: rules menu + research readers (experiments, docs, trials ledger)

[eaitbrahim]

Per the PRD §5 C4 + O5: the Rules menu (list/select, backtest, promote [confirm; --force typed], disable, enable, demote, add, insights) and the Research readers — experiments corpus, research docs, promotion reports, trials ledger list/verify — browsable in-console. Acceptance: all actions dispatch to C1 services; readers render the actual files/ledger with no re-parsing logic of their own beyond presentation.

[eaitbrahim]

Scope addition (2026-08-19, operator): the strategy console (PRD O11). The Rules menu becomes the full try/evaluate/retry loop, all through the C1 services: (1) run simulations from the console and display results — run_simulation on the active profile, GO-LIVE/TRAIN-MORE verdict, gates + numbers, DCA benchmark, tier matrix, written report linked from the Research readers; (2) the tried-vs-used ledger — every rule with its lifecycle status AND the machine's recorded reason (the specific failing promotion floor, insights distance, disabled/demoted context) — the engine's own verdicts, never a TUI narrative; (3) add a new strategy — the rules add flow in-console with O8 parameter help at each field, landing as candidate exactly as the CLI does; (4) retry a strategy — re-backtest + re-attempt promote (confirmations per O3, --force typed) and rules enable as the documented restore path.

[eaitbrahim]

C4 complete — merged in PR #404 (adversarially reviewed: one BLOCKER caught and redesigned — the ledger originally re-backtested every rule on entry, and rsi_meanrev is intrinsically quadratic (448s on a 5y hourly window; ~2.4h on the 19-rule hourly deployment, uncancellable). The redesign: the ledger renders only RECORDED state at entry (rows, stamps, insights' kind-wide paper distance — zero backtests, spy-proven on a 19-rule fixture); the full-window verdict re-compute is an explicit per-rule Enter action with a minutes-on-long-series warning, honest blocking like simulate/fetch, held-verdict invalidation on re-entry. Plus: two byte-compat breaks fixed (rules_disable swallowed its stderr error; promote loaded config before the unknown-id refusal), poisoned rows no longer blank the ledger, five 80-col overruns rewrapped, pullback's dead granularity form-field filtered to persisted params, demotion context rendered with honest column wording, midnight report-path drift and discarded-progress-on-failure fixed, kind-wide distance disclosed on the rendered line).

The strategy console: tried-vs-used ledger (the engine's own verdicts — can_promote's specific failing floors verbatim, 'no verdict computed' when absent), simulate with an ARMED plan (target report path shown first, Enter confirmed, verdict/report rendered verbatim under a pinned footer), add-a-strategy with per-field help from describe_params (introspected from the four rule classes' new PARAM_DOCS — the O8 single source), retry (re-backtest + promote with confirm; --force behind the CLI's own typed gate), enable as the documented restore. Research readers: experiments/research corpora and promotion reports (bounded 1 MiB, mtime-cached), the trials ledger with read-only verify. Extractions to the service layer: add_rule_row, run_rule_backtest, attempt_promotion, apply_rule_enable/disable/demote — CLI byte-compatible (17 scenarios diffed vs origin). A CI-only test failure (wrap-boundary phrase assertion under long runner paths) was caught by the full checks and fixed wrap-normalized. Gates: 3470 passed, 3 skipped; ruff/mypy clean.