Protected record delivery¶
Curated source-specific evidence, not a new run or qualification of current source. Dates, measurements, failure outcomes and original source identifiers retain their recorded meaning. Old tags/assets have been retired and original commits may no longer resolve after the privacy rewrite. Where identifiers were sanitized earlier, the original hardware/run association cannot be independently recovered from the public derivative. Current acceptance lives in the roadmap.
Recorded 2026-10-12; qualified source
180782fcd03702ff25014578f6b2432f527ccb2a. Hosted runs 38189341603 and
38189341363 passed for that source. Later correction results are retained in
data; original interval data is in
the baseline record.
Implemented boundaries¶
Exact Decoder::publish bytes pass through a separate bounded private record
socket into a C++ owner in the existing gateway. Default protected payload budget
128 MiB aggregate, bounded indexes/bootstrap and fair per-stream byte/entry/epoch
shares; decoder-worker/service restarts retain history, gateway restart destroys
it. Protected outbound copies survive eviction and no reader pins history.
Canonical envelope 1.0/configuration v2, accepted DSP and legacy browser/IQ/QoS
remain unchanged. Full contract/examples: record delivery.
Executed local evidence¶
Application implementation fdad9d2ad65cec2c4bdd74aa1a78e3309b82509c, with
requalified Origin correction at 3dfefe5cefebfddb5ad21ba32d3f924fbec88061.
Pi 5 ARM64, Debian Trixie, g++ 14.2.0, pinned Emscripten 4.0.23, Chromium
154.0.8037.92. Local browser profiles and loopback fixtures only; no remote CDP,
feature flags, TLS/security bypass or wasm32 substitution.
| Group | Executed result |
|---|---|
| Protected retention/framing | Native + ASan/UBSan/leak checks passed real locked mappings, DONTFORK, initial/replacement/lock denial, ownership/zeroing, concurrency/eviction, fair quotas, original bootstrap, exact u64/holes/conflicts and canonical migration. |
| Gateway/API | Real isolated normal and verified-TLS gateway passed independent consumers, body auth, privilege separation, exact Origin/path/listener checks, conflicts/retries, failed peak/CAS, inclusive ranges, large private/public fragmentation, invalid ACK/cursor/command recovery, replay/live races, slow-reader expiry, bootstrap, active revocation and owner-restart history loss. |
| Native service | Generated clean/corrected/rejected receive cases passed actual worker publication into gateway; earlier runs remained after decoder service stopped. Synthetic evidence is not independent RF qualification. |
| Actual browser memory64 | 154 delivery checks over 15 canonical fixtures, including 8 independent 892-byte header references; actual WS authentication/reconnect/native-byte/hash parsing additionally passed in a sandboxed local worker. Address type i64 and ABI pointers/lengths/u64 are verified. |
| Accepted receive regression | Complete native/sanitizer/actual local memory64 suite passed 32 documented weak-signal cases, frozen intermediate/pre-FEC/RS comparisons and independent OTA replay: 92 frames, 17 corrected symbols, exact SHA-256 below. |
| Legacy gateway | Existing HTTP/IPv4, TLS/IPv6, job auth/read-only/no-wake, concurrency/fault/handoff checks passed. Protected IQ queue regression also passed. |
Independent payload SHA-256:
70e147554060f5be9aa117737a6061fcc3f54f46accd58b3f9608862e8450677.
Delivery module (51,542 bytes, 16 MiB initial memory):
586d94230f0005b4dec72c203265f50b710e0120940b224a1e864181bd6f3894.
Small memory64 execution does not qualify actual addressing above 4 GiB; Phase 12
remains pending. Browser UA compatibility text says x86_64, while the executing
host/tool architecture is ARM64; it is not used as architecture proof.
API sanitizer integration instruments the gateway/records/core processing paths. ASan's process-lock interceptor/shadow space cannot qualify MCL_CURRENT for all memory; mapped pools use actual kernel lock wrappers. Normal service execution separately verifies process/TLS locking and non-dumpability. API LSan ptrace is incompatible with that protection and disabled only in this integration; core/ explicit ownership tests retain leak checks. No sanitizer finding is ignored.
A hostile-header regression initially reproduced oversized Origin acceptance:
lws_hdr_copy returns -1 on truncation, which was mistaken for native absence.
The record endpoint now checks the actual header length and fails closed on
truncation/copy failure. Truly absent native Origin still requires the same body
credential. This correction does not alter the legacy IQ/control origin policy.
Additional cross-run and same-run reversed-range regressions reproduced missing
start/end ordering validation. The endpoint now rejects those requests before
subscription admission, rather than accidentally accepting indefinite live-follow.
Valid inclusive ranges, supplemental evidence and replay fences are unchanged.
Hosted validate at 4cf3f23 passed native record/auth and instrumented gateway
tests, then failed legacy static-asset integration with HTTP 404: the API-only
plan had not built its manifest/catalog assets. Validation now explicitly includes
the web dependency for gateway integration, including direct --groups api runs;
no integration is skipped or converted to a successful no-op.
Reproducible commands¶
components/rfsentry-core/tools/test.sh
make -C components/rfsentry-api -j1 all test test-records test-records-sanitize
RECORDS_RECEIVE=1 python3 components/rfsentry-api/test/records_integration.py
RECORDS_BROWSER=1 python3 components/rfsentry-api/test/records_integration.py
./scripts/node-toolchain.sh run make -C components/rfsentry-api integration
# Explicit, metadata-only live measurement (never opens IQ/record subscriptions):
sudo ./scripts/node-toolchain.sh run node \
components/rfsentry-decoder/test/records-soak.mjs baseline satcom1-hrit \
components/rfsentry-core/build/phase4-evidence/baseline-live.json 600
# Use mode "records" after deployment, with separately launched record consumers.
Prepare core fixtures/module and compile build/record-fixture plus
../rfsentry-decoder/build/service-fixture before the optional modes. Test routes
are disabled; the harness uses only private synthetic traffic. Required hosted
core workflow now explicitly executes native-service, gateway/API/auth, ASan/
UBSan integrations and real browser authentication, rather than relying on a
configured/skipped group. General validation also executes the API integration.
the measured steady-state interval and has no invented exact lost-frame extent. Job revision 8 and configuration digest remain unchanged.
| Measurement | Accepted baseline | Phase 4, two record readers |
|---|---|---|
| UTC interval | 00:23:15–00:33:16 | 01:09:27–01:19:29 |
| Diagnostic elapsed seconds | 600.399 | 601.379 |
| Decoder CPU seconds/wall second | 0.466090 | 0.465763 |
| Gateway CPU seconds/wall second | 0.010510 | 0.086717 |
| Validated frames | 33,970 | 34,024 |
| Corrected symbols | 257 | 407 |
| Rejected frames | 0 | 0 |
| Source gaps / IQ overflow / output failures | 0 / 0 / 0 | 0 / 0 / 0 |
| Largest sampled last-block lag, ms | 20.642 | 20.379 |
| Largest sampled IQ queued bytes, 2-MiB limit | 246,528 | 0 |
| Worker RSS/locked/swap, KiB | 7,824 / 7,984 / 0 | 7,824 / 7,984 / 0 |
| Gateway RSS/locked/swap, KiB | 26,144 / 92,112 / 0 | 187,168 / 253,136 / 0 |
Decoder cost is effectively unchanged. Gateway adds 0.076208 core, including retention, TLS delivery, two independent readers, bootstrap and faster-than-live replay. Combined gateway/decoder adds 0.075880 core, about 1.90 percentage points of a four-core Pi's total CPU capacity; this excludes the two local test readers, rtl_tcp/demand and unrelated load. Reader CPU was not separately sampled. The later interval after catch-up still uses about 0.086236 gateway core, so replay startup alone does not explain away delivery cost.
No build/profile campaign or global CPU-policy change ran during either interval. Temperatures 60.9–64.2 °C / 60.9–63.7 °C, throttle flags 0x0 throughout. Sampled clocks were about 2.400 GHz; candidate had one 1.800-GHz governor sample, not a thermal-throttling flag. Load averages and reference readers are competing-load evidence. Lag/queue maxima are 30-second snapshots, not continuous percentiles/high-water proof. Different RF intervals/correction counts do not establish changed sensitivity or bit equivalence; the frozen-input native/ memory64 regressions establish unchanged decoding.
Independent delivery and memory evidence¶
Both protected, TLS-verifying C++ clients ran 660 seconds, ordinary forward
sequence 0–51047, 51,048 original identities. Every matched canonical
hash agrees; no forward gap/reordering/conflict. Reader A ACKs every 10 ms; reader
B every 200 ms, pauses 60–80 seconds, reconnects once at 180 seconds and resumes
after its original cursor. Its only duplicates are the two original supplemental
configuration/boundary records; forward cursor never advances from supplements.
Client traces contain hashes/identities only, in private /run tmpfs; no live
payload is persisted, captured or logged. Aggregate comparison:
small payload-free evidence.
Observed paced catch-up 4,800 records in 30.31 seconds ≈158.4 records/s while production is about 56.6/s. At 150.6 seconds it is still behind, with independent ACK/pending state; at the 179.9-second snapshot it has reconnected and both sent cursors equal newest 23994. Later snapshots remain at live. Maximum sampled sent-but-unacknowledged backlog 32 messages / 52,256 bytes, not an expanding queue or a retention pin. Cancellation/expiry/races/lost ACKs and restarts use small deterministic local/hosted fixtures rather than inventing live faults.
At soak end the actual ring held 47,831 records, 76,192,391 wire bytes, 85,710,592 rounded physical payload bytes, original bootstrap 794 bytes, one proven range 0–47830 and 845.417 seconds admission span. Default payload allocation 134,217,728 bytes, additional fixed indexes 27,262,976 bytes. Zero admission/partial-handoff/malformed/integrity/eviction losses during that interval. Capacity estimate ≈1,284 seconds is aggregate-production-based, conservative and not a promised outage duration, nor proof that RF never failed outside the measured interval. Fairness/expiry/gaps are separately tested. Continued live reception also reached the real default byte limit: 74,897 forward records, sequence 24613–99509, 24,613 normal evictions and 134,216,448 / 134,217,728 occupied/budget bytes, with zero admission loss. Original bootstrap remained 794 bytes and actual span 1,323.726 seconds. That is an observed bounded pool, not an outage-duration guarantee.
Kernel smaps shows lo dc dd (locked/no-fork/no-dump) on 128-MiB payload
and 26-MiB index mappings. Whole process/TLS heap/stack is locked via
MCL_CURRENT|MCL_FUTURE; dumpability 0, core limit 0, actual swap 0.
Service LimitMEMLOCK=1G is allowance, not allocation. Reference readers were
also locked, core-limited, zero-swap: 30,592 RSS / 31,136 locked KiB near end.
Effective memory-cgroup enforcement is absent; bounded provisioning and real
kernel locks, not a nonfunctional cgroup setting, establish protection.
The legacy compatible IQ client passed 1,471 frames / 96,521,136 bytes over IQ-only 8444 with zero sample/sequence gaps, no native decoder/capture restart, clean detach and owning-job rejection of incompatible open/tune.
A separate live revocation check authenticated one latest-follow reader, removed its credential through the operator API and observed active/allocated consumers 1 → 0 and expected client close (exit 1), without stopping reception. Both temporary test credentials are revoked; persistent policy revision 5 retains the same 128-MiB/default-age policy. Test credentials/traces are removed after aggregate validation; retained records remain untouched.