insights: one track record per rule; empty side reads unknown (#962, #965)

[eaitbrahim]

Closes #962
Closes #965

Rule track records on the Rules view, in keel insights summary and in its --json are now one record per rule. When one side of a sample is empty, its average now reads unknown instead of $0.00.

#962: one record per rule

Can trades be attributed to a rule? Yes.

• orders.rule_id is set on every paper entry by a real paper cycle, and each exit copies its entry's id.
• I checked the real paper databases read-only: every paper order in ~/keel/keel.db (dca and turtle_breakout) and in keel-paperhourly.db has a rule_id.

What changed:

• paper.track_records_by_rule(repo, kind) splits a kind's paper journal by orders.rule_id. The splits do not overlap, and between them they hold every trade of the kind (pinned). track_record(repo, kind) stays as the pool for the whole kind, because the journal's R match only knows the kind.
• Each record and its gate are built from that rule's own trades. RuleTrackRecord gains rule_id and product_id.
• A rule outside the scope keeps its trades. For example, a disabled rule's trades used to be pooled into the paper rule of the same kind. They no longer are.
• Trades journalled with no rule id (hand-built signals, or older fills) get one record per kind in scope, placed after the rules. That record has status="unattributed", no id, no product and no gate. They are not dropped, and they are not handed to a rule that happens to share the kind.
• Rules view. The records table has new id and product columns, and is still unsorted. Each gate card summary reads kind · product · rule N · verdict, and the card's id is h-gatedist-<rule id>. The position-based id that PR3 added is gone.
• CLI. Each record line reads [turtle_breakout] rule=10 product=BTC-USD status=…. The unattributed record prints rule=n/a product=n/a and a one-line note.

#965: an empty side is unknown, not $0.00

• Report. RuleTrackRecord.avg_win and avg_loss are None when that side has no trade. This is decided by counting the side's trades, not by reading the zero.
• Rules view. The payload's money(None) renders as "—" with state unknown.
• CLI. The CLI prints n/a.
• Gate verdicts are unchanged. BacktestResult, stats.py and promotion.py are untouched: pool_stats multiplies the zero by a zero count, and _realized_rr reads avg_loss_r. A test pins that each gate's (passing, blocking_reasons) equals check_floors on the same per-rule sample, and that the sample still holds 0 for the empty side.

Rulings

• One record per rule, not one per kind. The data attributes every real paper trade, and the ledger's "Follow-up 2" ruling chose per-rule when that holds. Cost if wrong: none, because a pool for the whole kind is still track_record.

• Gate cards now judge each rule's own sample, so the gate figures change. The verdicts here are display only. A kind-wide pool counted every product toward one rule's n_trades, without the Pool min_trades across same-parameter rules, with a cross-sectional diversity floor #338 diversity floor. keel rules promote decides from backtests plus pool_stats and is untouched. Cost if wrong: on the page, rules sit further from the floor than they did before.

• Unattributed trades get a record per kind, not a count and not a drop. This follows the Holdings precedent (Holdings: tranches group by rule NAME, so every DCA rule shares one "dca" table #960: unattributed tranches get one table per kind, placed last). Cost if wrong: one extra row, and only on a database that has such fills.

• --json shape: additive, plus null values.

  • Each record gains two keys, rule_id (int, or null) and product_id (str, or null).
  • avg_win and avg_loss are null when that side is empty. Before, they were "0".
  • status can be "unattributed".
  • No key was removed or renamed, and the gate object is unchanged.

  Cost if wrong: a script that does arithmetic on avg_loss must handle null.

• realized_rr is unknown when either side is empty (review round 1).

  • With no wins, the figure was gross win 0 over the losses, and that 0 is the zero insights: avg win / avg loss show $0.00 when that side of the sample is empty (unknown as zero) #965 calls unknown. The record and the gate card now hold None: the page shows "—", and the CLI says rr=n/a (no wins yet). With no losses it still says n/a (no losses yet).
  • The gate verdict is unchanged: check_floors reads promotion._realized_rr in R, and the engine's reason line still prints rr_r 0 < min_rr.
  • In --json, realized_rr (on the record and in gate) becomes null for an all-losses record. It used to be "0".
  • Cost if wrong: the gate card shows rr "—" next to the engine's rr_r 0 reason line.

• A zero-trade record reads unknown in the CLI text too (review round 1). The page already showed "—" for these (PR3 F5). The CLI line now reads win_rate=n/a avg_win=n/a avg_loss=n/a rr=n/a (no trades yet) expectancy=n/a profit_factor=n/a max_dd=n/a. --json keeps its numbers at n=0, so the JSON change stays additive. Cost if wrong: a JSON reader still sees 0.0 for win_rate at n=0, next to n_trades: 0.

• The client's record order is pinned (review round 1). A structural test checks that recordsSection hands records.map( straight to the table, with no sort, reverse or copy in between.

Tests and gates (head 534f134)

• Review. Three review rounds with mutation probes: round 1 had 21/23 killed (1 equivalent, 1 survivor, since pinned); round 2 had 18/19 (1 equivalent); round 3 had 9/9.
• TDD. Each test was red first: paper partition ×3, per-rule records, scope, unattributed ×3, CLI and --json identity, insights: avg win / avg loss show $0.00 when that side of the sample is empty (unknown as zero) #965 ×5 including the verdict pin, payload ×2, and view header↔cell pairs plus the gate card id and summary. Review fixes, each red first: all-losses realized rr unknown (record and gate card), the zero-trade CLI line (exact full-line equality), the client record order (structural), and realized rr counting losing trades.
• Static checks. ruff check keel tests packages, ruff format --check . and bare mypy are all clean (508 files). node --check passes on every js/*.js and on sw.js.
• pytest: 8329 passed, 3 skipped (head 534f134). logs/keel.log was deleted before the run (tests: suite leaves logs/keel.log in CWD; next run fails test_timeline_export header tests #964).
• Mutation. 12 probes, each proven applied and restored byte-identical. All are killed, and M4 was killed only after a pin was added for it:
  • kind-wide stats
  • empty side back to zero
  • unattributed dropped
  • unattributed folded into a rule
  • payload id dropped, or sent as ""
  • gate id built from the kind
  • id and product cells swapped
  • gate summary missing the product
  • partition key ignored
  • CLI printing None
  • CLI dropping the rule id
• Browser.
  • Setup. My own keel serve on port 8893, against a scratch copy of the repo's config.yaml and keel.db (eight turtle_breakout rules in scope). The copy was seeded with attributed paper trades plus one unattributed trade. Headless Chromium, with the service worker unregistered (0 registrations) and caches cleared.
  • Rules view.
    • 14 headers, each over its own cell, with 9 distinguishable records: ids 10, 11, 12, 21, 22, 23, 24, 25 and "—" (unattributed).
    • The all-wins record (rule 10) shows avg loss "—", and the all-losses record (rule 22) shows avg win "—".
    • 5 gate cards: turtle_breakout · BTC-USD · rule 10 · blocked, and likewise for 11, 12, 22 and 24.
  • Page behaviour.
    • 0 sort buttons.
    • 0 duplicate ids.
    • After opening card 11 and waiting one poll, only card 11 is open.
    • 0 console errors.
• CLI. keel insights summary (text and --json) on the same scratch copy shows rule=… and product=… on every line, avg_loss=n/a and avg_win=n/a on the empty sides, and null in the JSON.

🤖 Generated with Claude Code

[eaitbrahim]

Gatekeeper: ready for the owner to merge (head 534f134, base main 76382ee). I have not merged.

Review loop. Three rounds.

• Round 1 found 0 defects. I then promoted three held items to fixes:
  • realized rr reads unknown when there is no win;
  • the CLI prints n/a for every figure of a zero-trade record;
  • a structural pin keeps the client's record order.
• Round 2 found 2 defects:
  • This PR's description had been overwritten with Halt and keel serve hardening: drain every POST, HALT on a failed Pulse, serve keeps --external-host (#955, #958, #971) #977's. Merging it then would have closed server: /api/gates/halt leaves an unread request body on a kept-alive connection #955, Console: HALT should stay reachable when the Pulse payload fails to build #958, keel serve drops --external-host after binding: remote Host refused, session lifetime and release locality check hollow #971 and Console reads 'keel isn't running' after another process closes the WAL database last (setup.inspect mode=ro) #976 instead of Track records are keyed by rule kind: same-kind rules show identical, unidentifiable records #962 and insights: avg win / avg loss show $0.00 when that side of the sample is empty (unknown as zero) #965. I restored it from the edit history, and the closing references are again Track records are keyed by rule kind: same-kind rules show identical, unidentifiable records #962 and insights: avg win / avg loss show $0.00 when that side of the sample is empty (unknown as zero) #965.
  • realized rr decided the loss side by reading the average, not by counting trades.
• Round 3 found 0.

Evidence on 534f134

• CI: 4/4 green.
• pytest: 8329 passed, 3 skipped.
• ruff check, ruff format --check and bare mypy are clean. node --check passes on every JS file and sw.js.
• Mutants were killed in all three rounds; the only survivors are equivalent ones.

What I checked

• Promotion untouched. promotion.py, stats.py, agent.py and rules.py are unchanged. Each gate verdict equals check_floors on that rule's own sample.
• Browser. My own keel serve against a scratch DB, with the service worker unregistered, at 1280px:
  • 14 headers, each over its own cell, and 9 distinguishable records;
  • rule 22 (all losses) shows avg win "—" and realized rr "—";
  • 0 sort buttons, 0 duplicate ids, no page-level horizontal scroll (the table scrolls inside its wrapper, as it did on main);
  • 0 console errors.
• CLI. keel insights summary shows rule=/product= on every line, rr=n/a (no wins yet) / (no trades yet), and n/a for every figure at n=0. --json adds only rule_id and product_id; its new nulls are announced in the description.

Issues filed

• insights: an all-wins profit factor reads "—" (unknown) on Rules but "Infinity" in the CLI #979: an all-wins profit factor reads "—" on the page but "Infinity" in the CLI, and an all-scratch profit factor reads 0.00.
• promotion: check_floors' blocking reasons print win_rate 0.0 / rr_r 0 where the sample has no trades or no wins #980: check_floors' reason lines print win_rate 0.0 and rr_r 0 (engine wording).
• On tests: suite leaves logs/keel.log in CWD; next run fails test_timeline_export header tests #964, a comment: the stray log also breaks a test within a single run.