Upstream observation: session:start records no invocation provenance (argv/mode/isatty/launched_by) — human-vs-agent classification is heuristic-only

## Observation (vision-observation class; foundation-level, filed here per the #256 precedent — the foundation tracker is disabled)

Two independent forensic investigations of a real local corpus (46,715 sessions; CI-graph + local-file), run while building attractor-scout's session-provenance layer (PR #307), converged on this finding:

**No positive human marker exists in recorded session data.** The `session:start` payload carries only `{parent_id, redaction, session_id, timestamp}` — no `argv`, no `--mode`, no `stdin.isatty()`, no `launched_by_session_id`. A harness-fired `amplifier run --mode single` root session is byte-indistinguishable from a human typing at a keyboard: same event shapes, same `prompt:submit` events, empty `parent_id` by design.

## Why it matters
Any tool that mines a user's own session history (attractor-scout is the first; personalization/onboarding tooling generally will follow) cannot distinguish "work the human did" from "work an agent did on their behalf" except by heuristics. PR #307 ships a measured classification ladder (definitive agent signals: parent linkage, composite sid shape, first-event `session:fork`, orchestration events; then workspace/cadence heuristics) — but its named residual bypass is structural: **a harness that paces its prompts like a human in a stable workspace is indistinguishable in today's data.** The HOME-workspace one-shot band measured ~21–29% agent-contaminated and had to be classified UNKNOWN.

## The fix shape (one boolean collapses the whole heuristic tier)
Record at `session:start` (foundation/app-cli seam):
- `argv` / invocation mode (`interactive` vs `single`/batch)
- `stdin.isatty()`
- `launched_by_session_id` when a session is programmatically driven by another

Downstream consumers then get a definitive test instead of a ladder of measured heuristics.

## Provenance
Investigation and measured accuracies: `docs/designs/2026-08-20-scout-session-provenance.md` (this repo). Consumer that needs it today: `skills/attractor-scout/scripts/attractor_scout/provenance.py`.

Start with docs/designs/2026-08-20-scout-session-provenance.md and skills/attractor-scout/scripts/attractor_scout/provenance.py, then trace the foundation/app-cli seam where session:start is emitted. Review the existing payload and entry points; done means invocation context and programmatic launch provenance are recorded for downstream consumers.

Written by the indexing model from the issue text.