Summary

Closes #3143. Follow-up to #3142 — extends the same drift-aware re-render
pattern to the two other runtime-rendered prompt tails the boot-time
TOML drift loop was clobbering.

What was still broken after #3142

#3142 fixed ## User Configuration. The drift block
(kernel/mod.rs:~2929) was still stripping two more rendered tails
appended by activate_hand_with_id (~kernel/mod.rs:9131-9168) on
every restart:

• ## Reference Knowledge — skill content from def.skill_content
  (shared) or def.agent_skill_content (per-role override). Loss
  meant the agent dropped its skill discoverability summary; it
  could still call skill_read_file({skill, path}) on demand but only
  if it knew what to ask for, so first-turn skill consideration was
  silently degraded — most visible on session_mode=new triggers and
  cron fires.
• ## Your Team — peer roster for multi-agent hands. Loss meant the
  agent stopped knowing who its peers were and started calling
  agent_list redundantly at the top of every session, plus
  agent_send to peers became guesswork.

Both were silent: no error, no warning, just degraded context.

Approach (option B from #3143)

1. New idempotent helpers in manifest_helpers.rs modelled on
   apply_settings_block_to_manifest:

   • apply_skill_reference_block_to_manifest(manifest, role, def)
     — strips any pre-existing ## Reference Knowledge tail, then
     re-appends from per-role override (def.agent_skill_content)
     falling back to shared def.skill_content. When neither is set
     the call is strip-only — covers the case where a skill was
     removed from the hand and the stale tail must drop.
   • apply_team_block_to_manifest(manifest, role, def) — strips
     any pre-existing ## Your Team tail, then re-appends the peer
     roster excluding self. No-op for single-agent hands; gracefully
     drops the tail if a hand was edited from multi-agent down to
     single-agent.
   • Both use marker-anchored truncation
     (SKILL_REFERENCE_TAIL_MARKER, TEAM_TAIL_MARKER) for
     idempotency, same pattern as USER_CONFIG_TAIL_MARKER.

2. manifest_for_diff() projection strips all three known
   runtime-rendered tails before the JSON compare in the boot drift
   loop. Without it the disk TOML always looks "different" from the DB
   blob on hand-with-tails agents and drift fires every restart,
   burning a save_agent write and clobbering the rendered tails.
   With it, drift only fires when the raw TOML genuinely diverged
   from the raw DB form.

3. Activation site refactored to call both new helpers instead of
   inlining the rendering. The drift block calls them after the
   existing settings re-render. Both code paths now route through one
   source of truth — same pattern fix(kernel): re-render hand [[settings]] tail after boot-time TOML drift #3142 established for settings.

Why approach B and not A

#3143 outlined two paths. A (per-tail helper, no projection) only
prevents tail loss on a legitimate drift fire; B adds the projection
on top, so a hand with rendered tails no longer triggers a spurious
drift fire on every boot. The class of bug recurs every time someone
adds a new rendered tail; the marker registry approach scales —
register the marker, the projection picks it up automatically.

Tests

14 new unit tests + 2 boot drift integration tests, all green:

• apply_skill_reference_* — appends when present / no-op when empty /
  idempotent / replaces stale tail / drops tail when skill removed.
• apply_team_block_* — no-op for single-agent / appends peers
  excluding self / uses invoke_hint over description / idempotent.
• manifest_for_diff_* — strips all known tails / strips any subset /
  no-op on prompts without tails.
• boot_drift_preserves_skill_and_team_tails — multi-agent hand with
  SKILL.md, force drift via body-text edit, assert both tails are
  re-rendered after reboot.
• boot_drift_skipped_when_only_rendered_tails_differ — proves the
  sanitized projection prevents spurious drift fires when nothing
  changed except the rendered tails (which are runtime-only and
  expected to differ between disk TOML and DB blob).

cargo test -p librefang-kernel --lib
# 26 new tests pass; the 9 pre-existing failures
# (config::tests::test_load_config_defaults + 8 cron_delivery::*)
# are unchanged from main and unrelated to this PR.

cargo build --workspace --lib, cargo test --workspace, and
cargo clippy --workspace --all-targets -- -D warnings all clean.

Risk / blast radius

Low. Helpers are idempotent; the activation path's behaviour is
byte-identical to the old inline rendering on a fresh hand, and
distinguishable only on a re-render where the new path correctly
replaces a stale tail rather than letting it accumulate. Standalone
agents (no hand:<id> tag, no role tag) skip the new code path
entirely.