Ysopwnableandfactory-droid[bot]
23de4f77c7
Add explicit MCP Dart AOT summary delivery and exact retained string-pool pages while preserving complete defaults, CLI behavior, ABI selection, original Evidence identity, partial coverage and lower bounds. Real LocalSend SDK replies shrink 92.05% with approximately unchanged SDK time. Two actual Codex pairs preserve all selected facts and reduce total input 48.75% / 42.38%, uncached input 33.63% / 42.52%, and task time 23.95% / 30.30%; all client errors/repeated calls remain counted. Preserve old portable Evidence and the real CLI complete result after rebasing onto SAZ/qlog main. Curated results keep measured and integration runtime pins separate; raw artifacts stay outside Git. Refs #2016. Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
76 lines
5.3 KiB
Markdown
76 lines
5.3 KiB
Markdown
# Latest-main efficiency continuation
|
|
|
|
Continues [#2016](https://github.com/morluto/rea/issues/2016). The 2026-10-13
|
|
fetch resolved main to `f254f5ac2b804b353a711264bb178f43fdf07867`, the same
|
|
revision as the preceding Android conclusion. Newer capabilities below were
|
|
added during that earlier cycle; they are additional measurement scopes, not
|
|
new commits after that conclusion.
|
|
|
|
| Capability | Existing controls | Next evidence needed |
|
|
| ----------------------------- | ---------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- |
|
|
| Flutter / Dart AOT (#1582) | Explicit ABI, bounded string observations, partial coverage | Real shipped snapshot/count question; opt-in delivery only if unrelated strings dominate |
|
|
| SQLite (#2271) | Compact inventory, selected table/columns, row limit | Real application database and model workflow before adding another selector |
|
|
| TShark (#2318) | Fields, filter, decode-as, pass, default 50 matching rows, paging, explicit null for all | Real public capture and repeated task cost; paging repeats dissection |
|
|
| sigrok (#2339) | Decoder/channels/annotation selection, default 50 display lines, paging | Real public capture; display-line count is not annotation count and paging repeats decoding |
|
|
| Retained unknown head results | Existing retained paging | Reuse the existing view rather than duplicate its implementation |
|
|
|
|
Attached IDA, Frida and other live runtime workflows involve separate authority
|
|
and lifecycle questions and are outside this static parsing continuation.
|
|
Plugin SDKs and cross-session engine caches remain deferred until an observed
|
|
need and benefit justify them.
|
|
|
|
## First selected scope: Dart AOT
|
|
|
|
The official LocalSend 1.18.2 arm64 APK is pinned in `corpus.json`. The initial
|
|
real public SDK baseline inspected actual bytes without executing/installing
|
|
the APK. Its complete default AOT reply was 151,341 compact client-result JSON
|
|
bytes (70,487 normalized bytes), with 500 package URIs, 216 Dart URIs and 2,000
|
|
class-like tokens. String-pool and overall coverage were partial. A second
|
|
explicit arm64 request returned the same normalized facts.
|
|
|
|
These are exploratory first/warm calls, not an optimization pair. The frozen
|
|
runtime source is `28c84e78cf2fbfde4aceed345b773c59eb889f68`, with production
|
|
source bytes verified equal to main f254 at baseline acquisition. Its compiled
|
|
file-table digest is
|
|
`5332dda13981f043e146b3601ebec1f380d9f698280ca53d75786b638a912e11`;
|
|
the receipt is not relabeled as a newer build.
|
|
|
|
The candidate adds explicit MCP `detail: "summary"` and retained pool pages.
|
|
Omitted/complete requests and the CLI preserve the old complete path. Analysis,
|
|
ABI selection and original Evidence stay intact. Summary retains target/libapp
|
|
identity, exact nullable snapshot observations, lower bounds, returned counts,
|
|
coverage and limitations. Pages preserve original string order and source
|
|
coverage without rescanning. No engine cache, provider protocol or third-party
|
|
plugin API is introduced.
|
|
|
|
Follow-up measurements must compare identical selected facts, portable old
|
|
Evidence and actual model tasks. Reply bytes do not establish token or latency
|
|
improvements. Whole-scope final counters remain unknown if their cgroup has
|
|
disappeared, even when the narrower payload-exit observation is verified.
|
|
Raw records, binaries and model traces stay in the private lab outside Git.
|
|
Heavy work remains serial under the unchanged one-CPU, 3 GiB early-stop / 4 GiB
|
|
hard memory, no-swap and host-availability guards.
|
|
|
|
## Main advanced during the experiment
|
|
|
|
Main subsequently advanced to `a2d8b94dd50dab8d966239dd63ab84718aa8b4b4`,
|
|
adding native Fiddler SAZ parsing. Its web-network capture contract already
|
|
offers opt-in summary and retained capture pages, archive-member coordinates
|
|
and exact exposed encoded bytes. It decodes all records before ordinal
|
|
selection and rejects malformed/oversized inputs without partial success.
|
|
Measure real public SAZ artifacts using these controls before adding another
|
|
API or changing decoding or credential-redaction semantics. SAZ performance
|
|
remains unverified in this Flutter continuation.
|
|
|
|
The Flutter cohorts keep their original frozen runtime; subsequent integration
|
|
onto this newer main does not relabel earlier measurement receipts.
|
|
|
|
Main advanced again to `bb154d106` during CI, adding retained aioquic 1.5.0
|
|
qlog 0.3 JSON parsing and qlog views. The existing capture summary/page path
|
|
already applies. Qlog's trace-scoped STREAM interval unions and endpoint loss
|
|
declarations have separate semantics: they do not recover bodies, prove network
|
|
loss or establish PCAP correlation. Measure genuine supported recordings and
|
|
existing retained views before adding delivery or caching. Other qlog layouts
|
|
and qlog performance remain unverified in the Flutter experiment. Shared-view
|
|
integration is checked after rebasing; earlier runtime receipts remain pinned.
|