feat(library): six book-control rules for the risk category

[AetherAI3]

The risk/ category shipped one rule — max_drawdown_breaker, which is also the rule now sitting above the fold in the README, because it's the one that demonstrates host governance without implying a trade.

So the category that best sells the product had the least in it. This fills it.

What's added

Rule	Signal	Ceiling
daily_loss_limit	DAY_LOSS_PCT	>= 2
position_concentration_cap	MAX_POSITION_PCT	>= 25
correlation_cluster_guard	CLUSTER_EXPOSURE_PCT	>= 40
stale_data_halt	FEED_AGE_SEC	>= 30
leverage_ceiling	GROSS_LEVERAGE	>= 3
consecutive_loss_circuit	CONSECUTIVE_LOSSES	>= 4

All six emit only pause / observe and escalate to a named agent. None proposes a direction, so the category carries no performance or advice implication — and it stays inside baseline IR (AND-only thresholds over host-supplied series).

One design decision worth calling out

Every series is nonnegative and rises as the situation worsens, so each control is a >= against a ceiling. The same direction on every rule is what makes a stack of them readable at a glance — you never have to work out which way a given control points. The new Book-control signals table in the library README documents the feed contract for all six.

The headers carry the distinctions, not the thresholds

The threshold is visible in the source. What isn't visible is why each rule exists separately, so that's what the comment headers argue:

• daily_loss_limit vs max_drawdown_breaker — peak-to-trough over the book's life can sit perfectly quiet through a catastrophic single day, if the book was up beforehand.
• position_concentration_cap vs leverage_ceiling — distribution vs total size. A 1x-gross book entirely in one name passes the leverage limit cleanly.
• correlation_cluster_guard vs position_concentration_cap — one instrument, vs many instruments that are secretly the same bet.
• stale_data_halt vs event_liquidity_halt — can the book be traded, vs are the numbers describing it real at all. A liquid market with a dead feed passes that rule and fails this one.
• consecutive_loss_circuit vs daily_loss_limit — sequence vs magnitude. Four small losses can cost almost nothing and still be the clearest evidence available that the regime changed.

Verification

IR fixtures are generated from compile_to_dict, not hand-written, so the .nano/_ir.json pairs match byte-for-byte by construction rather than by care.

Five execution tests pin the behaviour that would otherwise rot silently:

• the >= boundary fires — a limit that only trips past its number lets a book sit exactly on the limit indefinitely
• stale_data_halt emits PAUSE before OBSERVE, because the ordered run log is the product
• concentration and leverage each stay silent on the other's failure, proving they aren't redundant
• a losing streak that resets never trips the circuit — a rule counting losses rather than runs would fire there

401 passed, 2 skipped (was 378). Library 26 → 32 strategies.

Not included

No new category, no grammar change, no performance claims. The corpus disclaimer — "a conformance corpus, not a performance claim, live signal service, or trading recommendation" — stands unchanged and now covers seven risk rules instead of one.