[Brian Krabach (bkrabach)]

Summary

Add real repair/save/resume regression coverage against the installed Foundation
dependency, with the dependencies needed to run it from a clean locked install.
No CLI production Python, repair policy, runtime context default or CI exclusion
is changed.

• Exercise the existing installed CLI repair and save helper bodies with real
  SimpleContextManager, CLI SessionStore/Foundation storage, and the actual
  resume initializer across four save/load cycles and backup rotation.
• Preserve original tool-result payloads, correlation IDs and metadata. Check
  healthy no-op repair, late results, and parallel batches with a truly missing
  result. Do not keep the old assertion that genuine late results are deleted.
• Declare context-simple from its public main branch in the dev group only.
  The lockfile records the tested commit
  db641363013ff906b656f49498a87ab01f5c506f for reproducible CI. Updating that
  dependency's lock resolution will follow main; the manifest is not pinned
  to that commit.
• Refresh only Foundation's locked source to merged PR446 commit
  a5f91fc4765dd3c6a1496b400c4fe428e5aa06f2. Keep Core 1.6.1 and every unrelated
  package resolution unchanged. The only added package is context-simple.
• Record the clean locked-install requirement in existing test guidance.

Why the dependency repair matters

The previous cross-repo DTU run explicitly installed extra dependencies and
proved the regression behavior, but did not prove this repository's standard
locked setup. Its old lock still selected pre-fix Foundation, and the new real
context import was undeclared. A main-ref requirement alone does not advance an
existing Git lock. This PR repairs both gaps rather than skipping the tests,
injecting a neighboring checkout into sys.path, or adding a runtime module
dependency.

Upstream microsoft/amplifier-foundation#446 is already
merged and its six main CI jobs passed. That change owns receipt preservation:
it can move a genuine result next to its call, while preserving the whole
message; it does not promise immutable chronology. Ambiguous associations retain
their evidence and remain broken. This CLI PR adds regression coverage, not a
different repair algorithm.

Verification

After switching the dev dependency to @main, a fresh workspace-local Python
3.12.3 environment was created only through
uv sync --locked --all-extras --dev, with no supplemental dependency installs:

• Installed direct_url.json verified requested_revision=main and the exact
  locked Foundation and context-simple commits. Foundation/context imports came
  from the environment's site-packages, not neighboring checkouts. Core remained
  1.6.1. Only requested-ref metadata changed in this follow-up; the resolved
  context-simple commit and every other dependency stayed unchanged.
• Four targeted repair/session/store/root-state files: 64 passed.
• Full default CLI suite: 2,743 passed, 3 existing skips, 13 configured
  integration deselections, 1 existing xfail.
• Separate POSIX integration selection: 13 passed.
• uv lock --check --offline, locked sync check and uv pip check passed.
  The lockfile did not change during installation or test execution.

The original dependency-repair full run had six failures because the isolated
TMPDIR made existing Unix handoff socket paths 128 bytes, above Foundation's
explicit 103-byte bound. Only the private workspace TMPDIR was shortened
(endpoint 96 bytes); all seven handoff tests and the entire default suite then
passed without source/test changes, new skips or guard changes. The original
failure is retained. This main-ref follow-up reused the corrected private
TMPDIR, without changing product or test code.

Offline test controls positively exercised local AF_UNIX IPC and rejected IPv4
and DNS before tests. All external network access remained denied during the
checks. No live model, vendor or desktop qualification is claimed.

Earlier isolated cross-repo DTU validation independently ran the same committed
regression with Core 2.0.2 (64 targeted passes). The dependency-only repair and
main-ref follow-up use the actual committed Core 1.6.1 resolution. No new DTU or
Gitea was launched for these packaging increments. Local Python 3.11/macOS/Windows
were not executed; hosted matrix and POSIX integration checks remain the PR
gates. The previous head's eight hosted CI jobs passed; new-head CI must be
evaluated separately after this update.

Review and limits

The prior regression/design was reviewed through second-opinion by Astra and
Fable on supplied code/evidence. Fable's dependency-resolution condition is
addressed by the fresh locked-install evidence above. A separate independent
source/lock/receipt review approved the dependency repair for publication;
these are not GitHub approvals or full live-application qualification.

The test-only AST helper extracts known definitions from trusted installed CLI
source, never user/transcript input; it carries a statement-local S102 rationale.
Construction is mocked and root_state=None for the repair/resume cases: this is
not a full interactive REPL or shared-root-lock acceptance claim. No public
fixture corpus or raw machine/session evidence is added.

No merge, normal user installation, global settings edit, Legacy migration, or
change to the separately open Managed/Chat Completions PRs is included.