fix: reject named_session entries referencing pool agents (#574)

[rileywhite]

Summary

Adds a pool-agent check in ValidateNamedSessions (internal/config/config.go) that
rejects [[named_session]] entries where the referenced template has namepool/namepool_names
set or max_active_sessions != 1. This enforces the invariant documented in
engdocs/design/named-configured-sessions.md Section 8 at boot time.

• max_active_sessions = 1 is explicitly allowed (singleton cap, used by gastown pack)
• nil max (basic template, no pool fields) is accepted
• Namepool, max > 1, max = 0, and max = -1 (unlimited) are all rejected

Closes #574
Closes #504

What was tested

• make fmt-check — pass (0 issues)
• make lint — pass (0 issues)
• make vet — pass
• go test -v -run TestValidateNamedSessions ./internal/config/ — all 8 subtests pass
• Verified gastown pack compatibility (all 5 named-session agents have max=1, pass validation)

Review outcome

Design council (5 perspectives) identified the correct approach. Adversarial review caught a
critical bug in the first cut: max_active_sessions = 1 was incorrectly rejected as "pool",
which would have broken gastown pack configs (mayor, deacon, boot, witness, refinery all use
max=1 + named_session). What-about review confirmed the same issue from a second angle
(control-dispatcher implicit agent). Fix: changed check to allow max=1, aligning with the
runtime's isMultiSessionCfgAgent definition and the API layer's isMultiSessionAgent.

Neutral judge verdict: CLEARED.

Residual note: the codebase has three separate "is pool" definitions that don't fully agree
(config validation, runtime isMultiSessionCfgAgent, API isMultiSessionAgent). This is
intentional — config-time treats nil max as "basic template" while runtime treats it as
"unlimited = multi-session". A future cleanup could extract named predicates.

Changes

File	Change
internal/config/config.go	+7 lines: pool-agent guard in ValidateNamedSessions
internal/config/session_sleep_test.go	+80 lines: 5 rejection + 2 acceptance test cases

No documentation changes needed — the invariant was already documented.
No breaking changes for in-tree configs.

🤖 Generated with Claude Code

[rileywhite]

Note: this PR also resolves #504 (per-agent drain policy). PR #652, which addressed #504 separately, has been closed as this is a superset of that work.