fix(session): honor config.session.metadata on session:start in the Rust kernel

[Brian Krabach (bkrabach)]

Parity bug fix, not a new feature

session.metadata is an already-specified, already-tested, already-used channel. It is
dead at runtime.

Layer	State
Contract	docs/specs/CONTRIBUTION_CHANNELS.md:76 — session.metadata – runtime metadata snapshots.
Spec + tests	tests/test_session_metadata.py — "CP-SM: Kernel reads config.session.metadata and includes it as optional 'metadata' key in event payloads. Pure passthrough." Green today.
Python kernel	python/amplifier_core/session.py:154-165 — implements it.
Rust bindings (what actually runs)	bindings/python/src/session.rs:389-392 — builds {session_id, parent_id}. Metadata dropped.
Pure-Rust kernel	crates/amplifier-core/src/session.rs:318-324 — same omission.
session:fork	python/amplifier_core/_session_init.py:274-291 — already correct (shared Python helper, both session classes delegate to it).

Production does from amplifier_core import AmplifierSession, which is RustSession
(python/amplifier_core/__init__.py:16). So one contract has three independent payload
constructions, and two of them disagreed with it — on session:start and session:resume
specifically, while session:fork quietly kept working.

Why the tests didn't catch it: tests/test_session_metadata.py:13 imports
from amplifier_core.session import AmplifierSession — the pure-Python class. The
metadata tests have been passing, green, on code that no runtime consumer executes. That
switchover blind spot is the actual defect; the missing merge is just its first casualty.

Dead-channel evidence

Measured on a real corpus on one machine (77,553 events.jsonl files under
~/.amplifier/projects/): across a 3,000-file random draw, 2,983 session:start
records were sampled and 0 carried any metadata key. Observed data key-sets:

988  ('parent_id', 'session_id', 'timestamp')
937  ('parent_id', 'redaction', 'session_id', 'timestamp')
689  ('parent_id',)
361  ('parent_id', 'timestamp')

amplifier-app-cli has been writing session.metadata for spawned children since
session_spawner.py:440-449 ("enables kernel CP-SM passthrough on session:start/fork").
It has been landing on session:fork and vanishing on session:start.

The change (~15 LOC of product code)

• crates/amplifier-core/src/session.rs — session_metadata_passthrough(), one shared
  reader for both Rust emit paths so they cannot drift apart again. Its emptiness guard
  mirrors the Python kernel's if session_metadata: truthiness test, so absent or empty
  metadata leaves the payload byte-identical to today.
• bindings/python/src/session.rs — cache session.metadata in new() alongside
  cached_session_id / cached_parent_id, merge it into pre_event_data.
• crates/amplifier-core/src/session.rs — same merge on the pure-Rust emit.

No new event. No new top-level payload key. No change to
crates/amplifier-core/src/events.rs.

Tests

• bindings/python/tests/test_session_metadata_rust.py (new, 8 tests) — the parallel
  suite against RustSession that the existing tests could not provide: start/resume with
  and without metadata, empty-metadata parity with the Python guard, verbatim nested
  passthrough, session_id/parent_id not displaced, and session:fork pinned.
• crates/amplifier-core/src/session.rs (4 new Rust tests) — the passthrough reader
  (configured / absent / {} / [] / "" / null) plus session:start emitted with and
  without metadata.

cargo test -p amplifier-core     473 passed; 0 failed  (lib)
                                  12 passed; 0 failed  (integration binaries)
                                  19 passed; 0 failed  (doc-tests)
cargo clippy -p amplifier-core -p amplifier-core-py -- -D warnings    clean
cargo fmt   -p amplifier-core -p amplifier-core-py --check            clean

uv run pytest tests/ bindings/python/tests/ -q -m "not slow"
                                1031 passed, 1 skipped, 6 deselected

CI that will run on this PR (.github/workflows/rust-core-ci.yml): rust-tests
(cargo test / check / fmt / clippy), node-tests, and python-tests on 3.11 / 3.12 / 3.13.
proto-check.yml is untouched by this diff (no .proto changes).

Back-compat

1. Metadata absent ⇒ unchanged. Asserted on both classes
   (test_session_metadata.py, test_session_metadata_rust.py) and in Rust
   (execute_omits_metadata_when_not_configured). Empty metadata is treated exactly like
   absent, matching the Python guard.
2. New kernel + old app ⇒ unchanged. The key only appears when config.session.metadata
   is non-empty, so callers that configure nothing get today's payload byte-for-byte.
3. No new top-level keys. Everything lands at data.metadata because metadata is not
   in hooks-logging's promoted-key list — no hooks-logging change is required or included.
4. Existing corpus untouched. Every already-written events.jsonl stays exactly as it
   is. This improves future data only.

Live proof

Built from this branch (maturin build --release), installed into a scratch venv together
with the companion app-cli change (microsoft/amplifier-app-cli#278), then one real
amplifier run --mode single "..." session. The persisted
~/.amplifier/projects/<slug>/sessions/<id>/events.jsonl session:start line, verbatim
apart from formatting:

{"ts":"...","lvl":"INFO","schema":{"name":"amplifier.log","ver":"1.0.0"},
 "event":"session:start","redaction":{"applied":true,"rules":["secrets","pii-basic"]},
 "session_id":"a0122cbb-6438-4014-99cb-c63fefef9985",
 "data":{"metadata":{"invocation":{"launched_by":"cli","launched_by_session_id":null,
 "mode":"single","schema":1,"stdin_isatty":false,"stdout_isatty":false}},"parent_id":null}}

Negative control — same app-cli, released amplifier-core==1.6.0 swapped back in, same
command:

{"ts":"...","lvl":"INFO","schema":{"name":"amplifier.log","ver":"1.0.0"},
 "event":"session:start","redaction":{"applied":true,"rules":["secrets","pii-basic"]},
 "session_id":"bbbbd7c4-8a1b-4b26-ba73-edfcbf9dc577","data":{"parent_id":null}}

Metadata configured on the mount plan, dropped by the kernel — the bug, reproduced on
demand. Note also that hooks-logging needed zero changes: metadata auto-nests under
data.

Notes for review

• The switchover blind spot (§ "Why the tests didn't catch it") is a general hazard for
  any kernel payload contract, not just this one. Fixing it broadly — e.g. parametrizing
  the payload suites across both classes — is out of scope here, but worth a decision.
• The pure-Rust emit fix is included deliberately. Fixing only the bindings path would leave
  the same trap armed for the next field added to session:start.

Co-Authored-By: Amplifier 240397093+microsoft-amplifier@users.noreply.github.com

[Brian Krabach (bkrabach)]

PARKED at the maintainer's request (2026-08-20) — amplifier-core cannot change right now. Work is complete and live-proven (see body); converted to draft to prevent accidental merge. Unpark order: this PR first, then amplifier-app-cli #278. Tracking: microsoft/amplifier-bundle-attractor#308.