Needs more testing but good enough for now. I get the same results with this as with `vllm serve`.

…ers)

LaTeX book covering transformer inference, KV caching, distributed
serving, Dynamo architecture, KV-aware routing, token blocks, network
codecs, fuzzing methodology, and 16 bugs found across 36 fuzz targets.

All bug references use ch10 canonical numbering (ai-dynamo#1-ai-dynamo#16), all cross-
references resolve, all arithmetic verified. Includes TikZ diagrams,
whynotbox pedagogical environments, and marginnotes throughout.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…iteups for 14 open bugs

- Reformat all 16 issue markdowns to match GitHub bug template (Describe the Bug,
  Steps to Reproduce, Expected/Actual Behavior, Environment, Additional Context)
- Add upstream/ directory with per-bug subfolders containing issue.md, pr.md, and
  discovery.md for each of the 16 bugs found via fuzzing
- Add fixes/ directory with annotated patch files and regression tests for all 14
  open bugs
- Create 14 minimal fix branches (fix/*-{bug#}) off upstream/main, each with a
  single commit touching only the affected file
- Verified all 14 bugs still present in upstream/main as of 50af343
- Confirmed 2 bugs already fixed upstream (ai-dynamo#15 RadixTree scoring, ai-dynamo#16 TwoPartCodec overflow)

…(R-B Slice 3)

New `ConditionalDisaggCoordinator` in `disagg/coordinator/coordinator.rs`
holds the unified per-request state map (`DashMap<String, Arc<CdRequest>>`)
and ports the prefill flow off `PrefillCoordinatorImpl` against the
canonical types defined in Slice 2.

Public API mirrors `PrefillCoordinatorImpl` so tests swap by
constructor only:
- `new` / `new_with_watchdog`
- `observer` / `observer_callback`
- `active_count` / `status_for` / `has_active_request`
- `commit_output_blocks`
- `cleanup_failed_request` (single canonical sink — uses
  `CdRequest::failed_g1_block_ids()` from Slice 2)
- `PrefillCoordinator` trait impl (`ensure_started`, `on_usaa`,
  `observe_forward`, `on_request_finished`)

Lifecycle watcher uses `spawn_lifecycle_watcher` (the helper
from Slice 1).  `on_request_finished`'s observer-residual
finalize-deferral is preserved verbatim, including the 10s
watchdog that prevents the `commit_output_blocks` ↔
`session.finalize()` race (learning #16).

Decode-role bits are present on `CdRequest` but unreachable
through this slice's code paths — Slice 4 wires them.

Coexistence (Slice 3 only): `PrefillCoordinatorImpl` stays alive
because production wiring (`prefill_leader.rs`, `init.rs`,
`unified.rs`) still references it; `cd_loopback`, `cd_decode_e2e`,
`cd_bidirectional_e2e`, and the unified suites also still
construct it.  Slice 5 cuts production over; Slice 7 deletes it.

Deviation: `ConditionalDecodeG2Observer::new()` widened from
private to `pub` so the new coordinator (sibling module) can
construct one.  No behavior change.

Tests:
- cd_prefill_e2e: harness swapped to ConditionalDisaggCoordinator;
  all 11 tests pass against the new coord.
- lib unit: 197 → 198 (new `size_of_is_nonzero` constructor probe).
- All other suites unchanged.
- Total CD-scope: 229 active tests passing.

/ai-dynamo#8/ai-dynamo#15/ai-dynamo#16)

Test-only additions for the seams the review flagged as untested.

ai-dynamo#7 _project_scale_to with a real apply outcome (4 cases):
  both components changed → full ScalingDecision; both equal current →
  None (PSM-equivalent no-change); single-component proposal → other count
  stays None; non-apply execute_action → None. Previously every adapter.tick
  test hit only the None/empty path, so a regression in the projection /
  no-change detection would have shipped silently.

ai-dynamo#8 _tick_input_to_context + FPM encoding:
  build a TickInput with traffic (incl kv_hit_rate), worker counts (incl
  scaling-in-progress flags), and a real ForwardPassMetrics; assert the
  PipelineContext.observations mapping and that the FPM bytes decode back
  (key format "<worker_id>/<dp_rank>", canonical encoder). This is the
  ingress glue where the add_observations P1 + the projection live.

ai-dynamo#15 registry mutation during an in-flight tick:
  suspend a PROPOSE plugin mid-gather (asyncio.Event), register a new
  plugin while suspended, release, and assert the late plugin did NOT join
  the in-flight stage (pre-tick snapshot) and the tick completed cleanly —
  then a fresh tick picks it up. Exercises the no-locks invariant that
  scheduler.py/server.py document but no test covered.

ai-dynamo#16 test_tick_diagnostics_extended scope note:
  clarify in the module docstring that plugin_overrides / reconcile_reasons
  / held_over_plugins have no production populator in this PR; these tests
  lock the dataclass contract (defaults / no shared-mutable aliasing), not
  live behavior.

835 planner tests pass (+6).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>