Browser reconstruction regressions¶
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.
Legacy/shared-core parity audit (2026-10-22)¶
Conclusion: the timestamp grouping defect is not isolated. The Phase 7 extraction did not preserve all enhanced legacy behavior. Native/memory64 agreement checked the same C++ implementation in two targets; it did not prove that implementation retained the legacy assembler's lifecycle/integration rules. This audit does not establish universal parity or claim every possible defect has now been found.
The audit compared packet/file assembly, Rice decoding and row placement, segmented imagery, integrity/navigation metadata, retained-product repair, timeouts, ownership/resource behavior and the product publication bridge. Production code, installed artifacts, receiver settings and live browser state were unchanged. No profiling, throughput test, soak or new CI run was performed.
Executed differential evidence¶
A private C++ diagnostic reused existing fixture builders and linked the
unchanged native reconstruction libraries. It exported the exact VCDU inputs
and native results. A bundled Node diagnostic fed those frames into the real
legacy HritLritAssembler and processLritProduct, then compared original
hashes, text, decoded samples, missing rows, DCS fields and product metadata.
Legacy group selection/state functions were extracted unchanged from
satcom.worker.ts to avoid copying/reimplementing their policy; its public
segmentedNativePng generated comparison pixels. This is a targeted diagnostic,
not execution of the entire legacy worker or a new memory64 qualification.
19 scenarios executed. Mismatches are the audit results, not passed qualification assertions:
| Area | Finding and evidence | Classification |
|---|---|---|
| Per-strip timestamps | Two valid strips differing only by 40 ms form one complete legacy mosaic and two incomplete native mosaics. Live retained 16-strip evidence independently confirms it. | Confirmed regression |
| Duplicate VCDU | Same VCDU under successive original record identities produces one legacy product and two native products. The legacy duplicate-frame guard is absent. Original-record replay dedup is a separate mechanism and remains required. | Confirmed regression |
| Duplicate first packet | Repeating the first packet with the same packet sequence produces one complete legacy file; native emits an unnecessary partial followed by the complete file. | Confirmed regression |
| Mosaic integrity | With all pixels present but a constituent carrying a frame-gap flag, legacy remains incomplete; native labels the mosaic complete. Native counts received segment numbers as complete and keeps first-strip loss fields instead of aggregating constituent quality. | Confirmed metadata/completeness regression, not proof of bad pixels |
| Navigation origin | Starting at strip 1 gives full-image line offset 1 in legacy and strip-local offset 0 in native. Native changes segment origin to zero without navigationAtImageOrigin()'s adjustment. |
Confirmed metadata regression; live overlay impact not separately measured |
| Partial-row recovery | Both locate a missing Rice row at exactly row 1 and preserve equal valid samples. Later nonconflicting good-strip replacement produces the same complete final pixels. | Preserved in tested cases |
| Missing-region appearance | Missing rows/canvas pixels use legacy 64-pixel cells at shades 42/58, versus native 16-pixel cells at 56/72. This accounts for unequal placeholder pixel hashes, not unequal received samples. | Confirmed visual parity difference |
| Inactive file | Legacy emits a retained abandoned partial on its existing 15-minute expiry; native has no equivalent expiry API/tick and keeps the assembly active until another boundary or limit. | Confirmed missing lifecycle behavior |
| Orphan continuation | Legacy emits bounded per-file orphan activity; native counts the orphan packet without that activity stream. Neither invents a missing file header. | Confirmed diagnostics/integration difference |
| Conflicting overlap | Native rejects a differing overlapping strip with integrity, reaching the browser's terminal failure path; legacy chooses a candidate using its strip-quality policy. |
Confirmed behavioral difference; do not blindly copy permissive replacement or weaken integrity. Design per-product conflict isolation |
| Normal products/codecs | Normal weather, independent deflate ZIP, NOAA DCSH, unknown originals, known libaec Rice/low-entropy vectors, horizontal segment pixels and a missing-packet text case agree on the compared original bytes/content. Archive presentation fills weather metadata in both paths. | Preserved in tested cases |
Additional code-path findings:
retainedSegmentSources()explicitly excludesrecordProvenanceproducts. The legacy path can rebuild mosaics from retained decoded strips after in-memory state is lost; the record path has no equivalent integration. This is directly relevant to repairing the retained full-disk observations.- Legacy mosaics link constituent product IDs and preserve aggregate integrity; native mosaic metadata only retains the first/last record lineage and lacks the original-strip link list used by those product repair/details paths.
- Legacy file/group idle expiry, per-file/packet Acquisition Activity and orphan progress are not supplied by the native product port. Aggregate record status and the Errors archive are present, but do not replace those views.
- Legacy stores sparse strips with limits of 64 groups / 512 MiB and uses coalesced image rendering plus a separate compact-product processing lane. Native allocates full canvases under 16 groups / 256 MiB and encodes PNGs inline during ingest/finish-batch. Bounds are intentional and must remain; the changed behavior needs explicit interleaving/resource correctness proof, not a blind constant increase. No numerical throughput inference is made.
- Canonical validation, original run/capture/configuration separation and source-independent identities are necessary new guarantees, not features to remove when restoring legacy behavior. Likewise the downstream color recipes, catalog, viewers, playback, exports and OPFS product archive remain shared.
Outcome and remaining limits¶
The audit's native/memory64 agreement was insufficient proof of legacy behavioral parity. Subsequent corrections restored bounded duplicate/lifecycle handling, product integrity/navigation, retained-input reconstruction and source recovery. Server Phase 7 was operator-accepted on 2026-10-31; this does not turn the 19 mismatches above into passed assertions or establish universal parity.
Other consequential failures were a seven-byte idle packet rejected by shared reconstruction, transient reader transport failures stopping recovery, and SIGBUS from concurrently replaced mapped index state. Preserve the corresponding packet, retry/cursor/ACK, storage ownership and recovery regressions. A retry may not invent progress, change original identity or relax integrity checks.
The prior sanitizer helper timeout remains unresolved as recorded in onboarding. Platform/iPad recovery, >4-GiB execution, sustained resources and the integrated multi-server matrix require separate proof. Current contracts are ARCHIVE.md, record-source operation and the Server roadmap.