Expose observed task resource targets - #461

Problem

st2 tasks --json identifies live task generations but does not expose a truthful resource-observation target. Downstream collectors would otherwise have to parse transient scope names, treat PIDs as identity, or maintain a second registry.

Goal

Expose a typed, fail-closed resource target for each observed runtime generation without turning PID, cgroup path, or process generation into task identity.

Decisions

• Advance task inventory to st2.task-inventory.v2 with linuxCgroupV2, darwinProcessTree, and bounded unavailable variants.
• Derive Linux targets from /proc/<pid>/cgroup; never parse systemd unit names.
• Fence exec targets with kernel start tokens.
• For PTYs, bracket one socket-backed pty stats --json snapshot per inventory with kernel start-token reads. Require a unique live row with matching session name, daemon PID, and creation time.
• Keep PTY generation IDs based on stable PID/creation evidence. Transient target availability cannot change identity.

Verification

• nix develop --command cargo test --lib task_inventory::tests — 13 passed.
• nix develop --command cargo test --test task_inventory_cli — 14 passed.
• Focused PID-reuse, socket-proof, one-stats-per-inventory, generation-stability, and real PTY replacement regressions pass.
• Full tests: 684 passed, 10 failed, 1 ignored; pristine current main fails the same 10 tests with 677 passed, 10 failed, 1 ignored.
• New requirements, ontology, spec, decision, and experiment records pass strict VRS shape checks.
• Rebased onto Preserve argv through systemd scope wrappers #455/current main; no review threads are open.

Pre-flip deviations

• Full clippy stops at the unchanged manual_is_multiple_of warning in crates/st2-resource-protocol/src/lib.rs:410. The same package-level command fails identically on pristine current main.
• The 10 full-test failures are identical to pristine current main and do not exercise task inventory or resource targets.

Complexity

The socket-backed PTY proof adds one bounded stats snapshot per inventory. The extra proof is necessary because registry-only PTY metadata cannot exclude PID reuse.

Concerns

The Darwin target is a best-effort process-tree root. st2 does not aggregate Darwin process resources. Linux cgroup-v2 is the first downstream accounting backend.

Friction & bottlenecks

The repository-wide baseline deviations are documented above. The focused inventory and real PTY scenarios are green.

Follow-ups

• Define exact terminal accounting and causal provenance beyond sampled task resources #454 tracks exact terminal accounting and richer provenance.

References

• Depends on Preserve argv through systemd scope wrappers #455
• Stack #462
• Define exact terminal accounting and causal provenance beyond sampled task resources #454