v0.7.11 — zero-inference fault-injection replay + newline classifier fix

Adds deterministic, zero-inference fault injection over recorded replay traces:
forge replay <sid> --fault <name> (repeatable) injects into a deep copy of the
trace — the recorded session and workspace stay untouched — and re-drives it
through the real Agent.send loop. Five scenarios ship: truncate_output,
malformed_burst, wrong_edit_anchor, force_compaction, and authority_violation.
The report covers recovery, false-completion, action/tool efficiency, observation
failures, loops, escalations, authority denials, completion rejections, and
context-token pressure; inapplicable and unknown injections are reported
explicitly rather than silently skipped.

Also rolls up the post-v0.7.10 fix that treats unquoted newlines as command
separators in the authority classifier, so a privileged later line can no longer
hide behind a benign first command (quoted multi-line data stays exempt).