refactor(commands): split AgentAccess god-trait into 15 domain sub-traits

[bug-ops]

Summary

AgentAccess (crates/zeph-commands/src/traits/agent.rs) had grown to 53 async
methods across 40+ unrelated slash-command domains in a single flat trait, with a
single ~1,800-line impl AgentAccess for Agent<C> block in
crates/zeph-core/src/agent/agent_access_impl.rs. Every command handler that needed
one method statically depended on the entire 53-method surface, and there was no
seam for adding a new command group without growing an already-huge trait file and
impl block further.

• Splits AgentAccess into 15 focused sub-traits (MemoryAccess, GraphAccess,
  ModelAccess, SkillAccess, PolicyAccess, SchedulerAccess, LspAccess,
  SessionControlAccess, McpAccess, OrchestrationAccess, SubagentAccess,
  IntegrationAccess, TrackingAccess, WorktreeAccess, MiscAccess), each
  declared in its own file under crates/zeph-commands/src/traits/.
• AgentAccess becomes an empty marker supertrait
  (pub trait AgentAccess: MemoryAccess + … + Send {}) with a blanket impl, so
  dyn AgentAccess and every existing ctx.agent.<method>() call site are
  unaffected — pure interface-segregation refactor, no behavior change.
• Agent<C>'s corresponding impl SubTrait blocks move out of the single
  agent_access_impl.rs into per-domain crates/zeph-core/src/agent/*_commands.rs
  files (5 new: memory_commands.rs, graph_commands.rs, skill_commands.rs,
  orchestration_commands.rs, misc_commands.rs; 10 existing files extended),
  mirroring the file layout the issue asked for.
• The shared delegate_cmd! macro moves to a new
  crates/zeph-core/src/agent/command_macros.rs, re-exported pub(crate) for the
  5 domain files that use it.
• NullAgent now implements the 15 sub-traits directly instead of one
  AgentAccess impl.
• A handful of call sites in src/ that invoke AgentAccess methods on a
  concrete Agent<C> (not through dyn AgentAccess) needed a specific sub-trait
  import — Rust's supertrait-method-resolution shortcut only auto-resolves
  through dyn Trait objects, not concrete-type dot-calls.

Went through 3 rounds of architecture review before implementation (plan approval,
then a significant-verdict critique that caught a real trait-name collision with
the existing SessionAccess trait plus a macro lexical-scoping issue, then a
further amendment to combine what was originally scoped as two follow-up PRs into
this single PR per explicit scope decision). Implementation-level adversarial
critique and independent test-coverage validation both approved with no blocking
findings; the developer caught and fixed one real bug along the way (a
#[cfg(feature = "scheduler")] module gate that would have silently broken the
blanket AgentAccess impl under default, non-scheduler features).

Closes #6488

Test plan

• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins — 14862 passed, 0 failures
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler"
• gitleaks protect --staged --no-banner --redact
• Relocated ~1390-line test module for the moved methods verified byte-identical against pre-move bodies (2 intentional UFCS-qualification diffs only)
• NullAgent's 15 sub-trait impls verified behaviorally identical to the original single impl
• Rebased onto current origin/main; only conflict was in CHANGELOG.md (kept both entries)