…s (review fix) (#3161)

High-severity review fix on top of #3126. The
`apply_cache_markers_system_and_3` algorithm reserved a fixed
`take = remaining.min(n)` budget and then blindly called
`stamp_block_with_marker` for each tail message. When a message's
content was an empty `ApiContent::Blocks(vec![])` — produced by
`convert_message` for assistant turns whose only content was a
Thinking block — the stamp silently no-op'd while the budget slot
was already consumed. The rolling window therefore shrunk from the
promised 2-3 message stamps to 1-2 in real conversations, and the
30-50% input-token savings claimed by the PR were not realised.

Algorithm change:
- Walk `api_messages` from tail → head and only count messages
  where a marker actually landed.
- Replace `stamp_block_with_marker` with `try_stamp_block_with_marker`
  returning `bool`; skipped (empty) messages do NOT decrement the
  budget, so the window keeps walking back until 2-3 markers are
  successfully placed.
- All-empty pathological case still terminates cleanly with zero
  stamps — no panic, no out-of-bounds.

Tests added in `anthropic.rs::tests`:
- `empty_blocks_message_does_not_consume_breakpoint`
- `total_cache_control_breakpoints_at_most_4_invariant`
- `rolling_window_when_all_messages_have_thinking_only_falls_back_gracefully`

All 27 anthropic driver tests pass; clippy clean;
`cargo check --workspace --lib` clean.

* feat(runtime): cache_hit_ratio metric + trajectory field (M2/2)

Builds on PR-1 (#3126, the system_and_3 stamping). Adds the
single-value observability metric the dashboard surfaces so users
can see at a glance whether prompt caching is paying off for their
workload. Pure observability — no behaviour change.

`cache_hit_ratio = cache_read / (cache_read + cache_creation)`,
in `[0.0, 1.0]`. Returns `0.0` when both numerators are zero
(caching not active for this turn / model doesn't support it).

Wire-up:
- `crates/librefang-runtime/src/agent_loop.rs`: new `cache_hit_ratio`
  helper alongside `accumulate_token_usage`. Per-turn `tracing::info!`
  on `librefang::cache` target with agent / hit_ratio / creation /
  read so existing log pipelines (Tempo, Grafana) pick it up
  without dashboard changes.
- `crates/librefang-kernel/src/trajectory/mod.rs`:
  `TrajectoryMetadata` grows `cache_hit_ratio: Option<f32>`
  (`#[serde(default, skip_serializing_if = "Option::is_none")]`)
  so legacy trajectories deserialise to None and replays don't
  need to know about caching to round-trip.

Tests (6 new):
- cache_hit_ratio_zero_when_no_caching (both 0)
- cache_hit_ratio_full_hit (read=100, creation=0 → 1.0)
- cache_hit_ratio_partial (read=70, creation=30 → 0.7)
- cache_hit_ratio_cold_start (read=0, creation=100 → 0.0)
- trajectory_metadata_cache_hit_ratio_serde_round_trip
- trajectory_metadata_cache_hit_ratio_legacy_compat (legacy JSON
  without the field deserialises to None, no panic)

Companion fix: `scheduler.rs:555` baseline
`clippy::manual_pattern_char_comparison` (`s.find(|c: char| c == '/'
|| c == '?' || c == '#')` → `s.find(['/', '?', '#'])`). Same one-line
fix is also queued in PR #3144 / #3147; whichever lands first the
others rebase cleanly (identical content).

Verification:
- cargo test -p librefang-runtime --lib agent_loop: 167 ok
- cargo test -p librefang-kernel --lib trajectory: 14 ok
- cargo clippy -p librefang-runtime -p librefang-kernel
  --all-targets -- -D warnings: clean

Stack: independent, base main. Closes the prompt-caching-system-and-3
plan — system_and_3 stamping (PR-1) plus this PR cover the full
"borrow hermes-agent's cache strategy" scope.

* refactor(types): consolidate cache_hit_ratio onto TokenUsage

Two parallel implementations of the same cache-ratio math (one in
agent_loop with f32/0.0-on-empty, one in trajectory with Option<f32>)
were drifting apart in semantics. Move the authoritative implementation
to TokenUsage::cache_hit_ratio() in librefang-types and have both
callers delegate:

- runtime: tracing log uses .unwrap_or(0.0) for the f32 field, with a
  comment explaining how to disambiguate "no caching" from "0% hit"
  via the creation/read totals.
- kernel: trajectory::compute_cache_hit_ratio() becomes a thin
  re-export; with_cache_hit_ratio builder doc clarified that the
  exporter never sees TokenUsage and call-site wiring is a follow-up.

Test coverage for the math is now centralized in librefang-types
(4 tests), runtime drops its 4 redundant tests, kernel keeps a smoke
test for the re-export plus the builder/serde/legacy-compat tests.