Improve investing decision quality: constraints, EV neutrality, eval metrics

[vingrad]

Quick-wins pass on decision quality for the investing pack, found while evaluating the finance planner's scoring path.

What changed

Per-goal risk constraints now bind sizing (internal/finance/constraints.go, new)
The investing pack validates drawdown_limit / risk_tolerance constraints and even warns "sizing will fall back to conservative defaults" when they're missing — but they were never read; the risk budget came solely from pack/policy config. EffectiveRiskBudget now tightens the budget per goal:

• a drawdown limit caps MaxPositionPct (a total loss of one position can't exceed the stated limit) and MaxPortfolioRiskPct (five consecutive stop-outs stay within it),
• conservative tolerance halves the Kelly fraction and concentration cap; aggressive scales the Kelly fraction only (hard caps stay as guardrails),
• unparseable constraints leave the base unchanged, and the move rationale names what bound.

EV no longer rewards volatility
With the default flat winProb=0.5 and winFrac = RewardRiskRatio * lossFrac, expected value reduced to 0.5*lossFrac*(ratio-1) — monotonically increasing in volatility, so every initial plan tilted toward the most volatile ticker. The EV component is now neutral (NeutralEVScore = 0.5) until a signal carries an actual probability hint. Relatedly, a signal hint now only informs the ticker it names (a valuation hint on ACME no longer lifted GLOBEX too); unnamed signals (e.g. macro) still apply to all candidates.

Backtest harness measures decision quality (internal/backtest)

• BrierScore: mean squared error of top-move confidence vs. realized outcome (label = positive forward return of the top thesis, falling back to the analyst kill label).
• Per-decision ForwardReturn: top thesis return from decision time to scenario end. Computed after the replay — no future data reaches a decision.
• NoiseRobustness: share of non-kill events that did not trigger a material replan.

These make later scoring changes (winProb priors from fundamentals, vol scaling, calibration) quantifiable against the shipped scenario.

On the shipped scenario

Backtest: ACME thesis lifecycle
  versions created : 1
  kill precision   : 1.00
  kill recall      : 1.00
  brier score      : 0.054 (lower is better; 0.25 == coin-flip confidence)
  noise robustness : 1.00 (share of non-kill events not reacted to)
  hypothetical pnl : +8.00% (illustrative, not a strategy return)

The scenario's "max 10% drawdown" constraint now shows up in the rationale: ...; risk budget per goal constraints: drawdown_limit 10%.

Testing

• Table-driven tests for EffectiveRiskBudget/parsePercent; planner tests for constraint-shrunk sizing, neutral flat-prior EV, and ticker-targeted hints; harness tests for the new metrics.
• go build ./... && go test ./... -race passes; gofmt/go vet clean.
• No wiring, policy, schema, or pack-descriptor changes; everything stays offline-deterministic.