Adversary stalls on large-context corroboration passes (stream watchdog) with no verdict

## Summary
When the fresh-context `adversary` sub-agent is dispatched to corroborate spec-level call signatures against a large implementation surface (many files, several thousand lines total when read in full), it can exceed the 600s stream watchdog before returning a verdict. The result is a stall — no review output — rather than a CLEAN/FINDING result.

## Where it bites
Integration / wiring / composition stories reference many source files at once (each subsystem plus the seam that joins them). A faithful fresh-context corroboration pass reads all of them, and full-file reads blow the watchdog budget.

## Observed workaround (works, but should be the default)
Scoping code references to signature-bearing line ranges (Read offset/limit, ~350 lines per file) plus an explicit "read each file once, reason once, do not loop" efficiency directive clears the watchdog reliably (passes complete in ~70–135s) while preserving fresh-context honesty.

## Proposed remediation (either)
1. Engine chunks/scopes the code-reference payload by default before dispatch.
2. The adversary dispatch template defaults to line-range-scoped references + the read-once/reason-once directive.

## Impact
Without a default scoping strategy, the standard dispatch path invites stalls precisely on the highest-risk stories (cross-subsystem integration), where thorough corroboration matters most.