ci: run A2 app and session suites

[aurascoper]

Prerequisite for #32's operator matrix. Owned by the D0/CI lane, so ci.yml stays with its named owner rather than growing #32 by a twenty-second unrelated file.

CI runs 5 of 9 declared test targets

Package.swift declares nine test targets. The deterministic step names five. Two of the omitted ones carry #32's load-bearing regressions:

Target	Discovered	In CI before this PR
NeuralComposeAppTests	116	no
DialecticSessionTests	5	no

That is 121 tests CI has never executed, including the app-level fail-closed runtime, runtime-presentation, cancellation, and headless runtime-factory behaviour #32 depends on. Both targets pass on this base with 0 failures.

This also accounts for most of the 417-vs-616 gap. swift test list discovers 580 tests on this base; CI's filtered subset executes 417.

BCILLMTests and WorldModelDemoTests stay unselected deliberately — MLX .eval() faults on a GPU-less runner, and the WorldModel demo needs exported CoreML artifacts.

The guard is the point

A --filter that matches nothing exits 0. Adding the targets without a guard would risk exactly the failure mode this repository keeps finding: a green step that ran nothing.

Two layers:

1. Sentinel classes, one per target — AppViewModelRuntimeFailClosedTests, RuntimeFactoryClaudeResolutionTests.
2. Discovered-count floors per target, because sentinels alone would still pass if the rest of a target vanished. Floors are 100 and 5, below the current 116 and 5, so ordinary test additions never trip them.

One correction to the proposed guard

The directive specified a third sentinel, RuntimeIdentityPresentationTests. It does not exist on this base — it is one of #32's own 21 files, and swift test list | grep -c returns 0 for it here.

Including it would have failed this PR's CI on the first run. It is omitted with a comment telling #32 to add it as a third sentinel when that file lands, so the guard stays true at every revision rather than being aspirational at this one.

Verification

Rehearsed locally against this exact base:

swift test list                                  580 discovered
  AppViewModelRuntimeFailClosedTests               3
  RuntimeFactoryClaudeResolutionTests              5
  RuntimeIdentityPresentationTests                 0   <- arrives with #32
  NeuralComposeAppTests                          116
  DialecticSessionTests                            5

swift test --filter NeuralComposeAppTests \
           --filter DialecticSessionTests       121 executed, 0 failures

yaml parse                                       OK
git diff --check                                 clean

After this merges

Merge-forward docs/eeg-methods-scope into #32 again, confirm A–G plus 9eabad6 plus the new merge-forward all remain ancestors, confirm the #32 diff is still the same 21 application files, and confirm all three surfaces green — deterministic subset, the new A2 app/session step, and Python 53 + 89/89. Only then start the packaged operator matrix.