Files
2026-11-10 16:56:31 -05:00

96 lines
5.3 KiB
Markdown

# Native empty-keychain control — negative retention result
The plain C control completed **200 private keychain create/delete/CFRelease
operations successfully**, with `Security::AppleDatabase` heap inventory staying
at **1 object / 768 shallow bytes** before, after the first operation, after 200,
and after a five-second idle. This **does not reproduce** the retention seen in
private-PFX and SDK configured-store loading. Empty native create/delete/release
alone is therefore insufficient to explain that result on this machine.
This does not rule out deletion/resource interactions for populated imported
keychains, identify the retaining owner, or assign an SDK, CLR or Apple defect.
Private-key/certificate materialization and its associated native interactions
remain distinct. No production mitigation, resource-budget decision, G4 pass or
#799 closure follows from this control.
## Reviewed execution and isolation
Astra reviewed the actual C/Python source before execution and afterward
independently verified all 207 JSONL rows, 200 ordered operations/status triples,
reference balance, default/search-list fingerprints, all four native heap
counts/bytes, FD/file counts, source/runner/binary hashes, cleanup and warnings.
Only **one actual native execution** ran, serialized after the coordinator's other
ownership control. Source is `PureAppleControl.c`; collector is `pure_apple.py`.
The child is a plain native C program linked only to Security.framework,
CoreFoundation.framework and libSystem. It does not load certificates/PFX, the
OPC UA SDK or CLR. `SecKeychainCreate` used a fresh random filename beneath its
private mode-0700 task root, a generated password and `promptUser=false`. A returned
keychain's path had to exactly match that generated private filename/backing file
before deletion was allowed. Each returned create reference received exactly
one `CFRelease`, including failure paths. The program made no default/search-list
setter or restore call. Read-only `SecKeychainCopyDefault`,
`SecKeychainCopySearchList` and `SecKeychainGetPath` produced hashed before/after
fingerprints; owned Copy results were released, borrowed array elements were not
released separately. Existing keychain paths/passwords were never logged.
Native checkpoints were before, first, 200 and five-second idle. Each parent
snapshot used `heap --noContent -s` and numeric lsof FDs, with Mach task resident
bytes reported by the child. It did not invoke GC or allocator-pressure APIs.
The original fixed 180-second checkpoint/child limit, 20-second heap and
five-second lsof limits, bounded final wait and terminate/kill reap remain in the
collector; those probe/reap bounds can extend total wall time beyond 180 seconds.
Actual child monotonic duration was **8.995 seconds**; parent wall time including
probes was **9.288 seconds**. The child exited zero and was reaped.
## Actual results
| Checkpoint | AppleDatabase count / shallow bytes | Numeric FDs | Owned keychain files / FDs | RSS bytes |
|---|---:|---:|---:|---:|
| Before | 1 / 768 | 3 | 0 / 0 | 7,372,800 |
| First | 1 / 768 | 3 | 0 / 0 | 8,028,160 |
| 200 | 1 / 768 | 3 | 0 / 0 | 8,699,904 |
| Five-second idle | 1 / 768 | 3 | 0 / 0 | 8,699,904 |
All **200 create/path/delete OSStatus triples were `(0, 0, 0)`**. Each create
returned an owned reference, its private path was verified, and all **200/200**
references were released exactly once. All read-only default/path/search-list
query statuses were zero. Before/after fingerprints were identical, with search
list count **2** at both observations. Four query-owned Copy references were
released across the two fingerprint collections. No list setter was invoked.
Default fingerprint:
`08c2c77dd8eaf8042b4020991ea06c6215cc391532869bbf6ced2cf1126c5e63`.
Search-list fingerprint:
`c05fae700986dc7a6a147a938ef2adc4a274173420e137929ed58ad986eb2a3c`.
Owned keychain files were already zero before harness cleanup. The fresh owned
fixture root was then removed. Native class count is a heap classification, not
an open-handle count or an exclusive attribution of the modest RSS change.
The first attempt stopped at **compilation**, before a native child or API call:
strict C11 headers did not declare `explicit_bzero`. Replacing only that
caller-owned buffer wipe with a volatile-byte wipe made the build succeed; empty
fixture creation was moved after successful compilation. The failed build is
retained separately. Both compile attempts emitted **five unsuppressed
deprecated-public-API warnings** for the intentionally exercised legacy keychain
APIs. This is not a zero-warning build and no warning suppression was added.
## Evidence and reproduction
Selected, home-path-redacted receipts are under
`evidence/2026-11-10-pure-apple/`; `verification.json` retains the unchanged local
raw receipt hashes. Actual raw output is
`/tmp/opcua2-pure-apple-20261110b`; the first compile-only failure is
`/tmp/opcua2-pure-apple-20261110a`. Raw native snapshots stay private/local.
```sh
python3 scripts/validation/opcua-lifecycle-memory/native-retention/pure_apple.py \
--output /tmp/opcua2-pure-apple-FRESH
```
The output must be fresh. Fixed counts are not configurable; all generated paths
are task-owned and no user/fleet PKI is used. This negative control refines the
[A/B directory-load findings](README.md), and does not approve another experiment,
a native workaround, a public upstream post or a long run.