Repository navigation
chore: publication cleanup — strip review residue, correct five stale claims, and say what was designed rather than built - #20
Merged
Conversation
Doc comments referenced an internal task plan and review-finding ids that are not part of this repository. Rewritten in place: pointers now name the class, file or measurement they meant.
The docs carried finding IDs, review-round numbering and sprint-plan task references that no reader of this repo can resolve. Every one of them was attached to a claim that stands on its own, so the claim stays and the provenance goes: "because review finding R0-F8.3 refuted that" becomes the reason itself, and "Task 1.8 measured" becomes the device the measurement came from.
Same treatment as the docs, applied to the dartdoc in lib/: finding IDs and "the reviewer" attributions are replaced by the evidence they pointed at, so a comment explaining why the code is shaped this way now explains it rather than citing a ledger the reader cannot open.
…ked for them Four groups in diagnose_screen_test.dart were named "R0-F1:" and friends, which told a reader which review round produced them and nothing about what they assert. They are now named after the behaviour under test. The same substitution runs through the rest of the suite's comments.
no_fake_in_production_test.dart, mic_capture_test.dart and agent_loop_test.dart all argue from the times they were wrong, and that argument is the point — but it was told in finding IDs. Each defect is now named by what it was: the wrapper that laundered a reference, the conditional import whose second URI was never read, the scripted InferenceHost that answered through the real engine.
Every remaining "Task N.M" pointed at a plan no reader of this repo has. Each one was standing in for something concrete — the retrieval router, the pinned recogniser, the provisioner — so the concrete thing is named instead. Generated sources regenerated so the codegen freshness gate still reports clean.
integration_test/ was the last directory still citing sprint-plan task numbers and a review finding. The device suites are the ones a reader is most likely to open first — they are where the measured figures live — so the same rule applies: name the thing that was measured, not the task that measured it.
Four claims had gone stale, and they matter more than usual because these docs argue from measurement: - the noRetry table covered eleven sites; there are fourteen. The three speech ones are measured here for the first time and all three survive, so the split is six bound / eight unbound. - the golden figures were taken before form_autofill and before the second tool joined the registry, which widened every prompt: 933/14 chars is now 1283/15, the widest is 2713 not 2363, and "five of six" goldens carrying an em dash is now all seven. - TC-VOICE-FILL-01 said three different things in three files. It ran once, failed on its own fixture, was fixed, and was never re-run — which is what all three now say, README included. - demo-screen.md narrated an unmerged PR that merged long ago. Also corrects three README claims: no network is required at *inference* time, not "at any point after install"; not every figure names its device; and the mutation harness lives outside this repo, so its survivor counts are reported rather than reproducible from a clone.
… language with them Both interfaces, both fakes and both providers had zero production consumers: vision/OCR was a stretch idea that was never started, and telemetry never got a backend. Dead ambition reads worse than absent ambition, so they go. lib/engines/providers.dart also still called the app a skeleton and streamed 'FieldOps Copilot skeleton is running on the fake engine' — from a file whose two remaining providers are a deliberate host-test default, not an unfinished binding. It now says which it is, and why the device bindings live elsewhere.
…sidue analysis_options.yaml was the stock template — every rule commented out. It now turns on the analyzer's three strict modes and a chosen rule set, each with the reason it is there. The strict modes pass with zero findings, which is the useful part: an app that crosses two FFI boundaries and a SendPort had no implicit dynamic leaking through lib/. Two rules are off with a stated reason (comment_references and avoid_slow_async_io — the latter is backwards for an app whose argument is that the UI isolate must not block), and test/ and integration_test/ carry their own options file relaxing three production-lifecycle rules that misfire on a test double. lib/ keeps all three. Also: the Apple DEVELOPMENT_TEAM id is blanked at all three sites, the Android label matches iOS's "Field Ops Copilot", the two Flutter-template TODOs in build.gradle.kts are replaced by the actual reasoning, and pubspec.yaml loses the generated boilerplate and the dead fonts block.
…it covered The README sent readers here for tiers, mutation testing and the live-STT suite, and the page had none of the three. It now opens with what each tier cannot answer — which is why the next one exists — and the mutation section states the method, the three rules that make a result quotable, and every recorded run including the one with eight survivors, because eight published survivors say more than a table of zeros. The CI story moves here too: the Flutter pin (a freshness gate that fires on Google's release is a gate people learn to ignore) and the codegen gate, including why 'git diff --exit-code' alone is not enough and why a warm build_runner cache makes a local pass prove nothing. Also fixes the five dangling 'see above' references left from splitting the monolith, adds the three integration files the list had never mentioned, and corrects stt_test.dart from 'written and unrun' to the device pass that speech-to-text.md had already recorded. 'Tier 0/1/2' is gone everywhere — it was sprint-plan vocabulary that meant nothing to a reader here.
The distance between a working demo and a deployable product is mostly the decisions this section names, and a README that omits them is quietly claiming to have closed it. Eight of them, each marked as design rather than code. Key management leads, because it is the first thing an enterprise-mobile reader will ask and the honest answer is unflattering: the cipher is real and the key is a constant named 'not-a-secret', which protects a stolen file and not a stolen device. The Keychain/Keystore answer already has its seam — databaseEncryptionKeyProvider — and so does the credential-delivery equivalent one layer out. Naming the seam is what makes the design answer checkable rather than a wish.
Four passages spent more words on the history of the sentence than on the fact it states — a guard auditing the guard that audits it, a count corrected by explaining how it was miscounted. Past the third instance that reads as house style rather than as rigour, which undercuts the instances where it earns its place. The findings stay; the recursion goes. Also drops the last commit SHA cited from a doc.
The run has happened. All four places that said it never had — the test's own doc, voice-and-work-order.md, device-test-scenarios.md and the README status footnote — now record the pass, and the README row drops its asterisk. The first-attempt failure stays where it is useful, because what it demonstrated is the reason the device tier exists: the fixture delivered nothing until asked, so the screen reached 'listening' after the assertion rather than before, and the app was right while the double was wrong. A double gentler than the hardware is a test that cannot fail. Also renames speech-to-text.md's 'Owed on a device' heading, which had said owed since the run it records passing.
karolkulesza
added a commit
that referenced
this pull request
Aug 20, 2026
…references chore: publication cleanup — strip review residue, correct five stale claims, and say what was designed rather than built
karolkulesza
added a commit
that referenced
this pull request
Aug 22, 2026
…references chore: publication cleanup — strip review residue, correct five stale claims, and say what was designed rather than built
karolkulesza
added a commit
that referenced
this pull request
Aug 23, 2026
…references chore: publication cleanup — strip review residue, correct five stale claims, and say what was designed rather than built
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Applies the publication review, points 1 and 3–7. Point 2 (rewriting
Claude-Session:trailers out of history) is left alone as recommended.1 · Strip the private-reference layer
~900 references across
lib/,test/,integration_test/anddocs/to things no reader of this repo can resolve: finding IDs (R0-F1,M19), "the reviewer", "review round N", andTask N.Mpointing at a sprint plan that does not ship here.The rule throughout was keep the claim, delete the provenance. Where a comment said "…because review finding R0-F8.3 refuted that", it now says why. Where it said "Task 1.8 measured…", it names the device. The hardest case was
no_fake_in_production_test.dart, whose entire argument is the nine times it was wrong — each of those is now named by what it was (the wrapper that laundered a reference, the conditional import whose second URI was never read, the scriptedInferenceHostthat answered through the real engine) rather than by a ledger id.TC-XXX-NNrequirement ids are kept — they read as a traceability matrix.3 · Doc staleness — five contradictions, all verified against the code
noRetrysites. The table covered eleven; there are fourteen. The three speech-path sites were measured here for the first time (delete the policy, run the whole suite) and all three survive, so the split is six bound / eight unbound.form_autofilland before the second tool joined the registry, which widened every prompt. 933 chars/14 lines → 1283/15; widest 2363 → 2713; "five of six goldens carry an em dash" → all seven (no_manual_matchgained one from the work-order tool's instruction).listeningafter the assertion rather than before, and the app was right while the double was wrong.stt_test.dartwas "written and unrun" intesting.mdwhilespeech-to-text.mdrecorded it passing 3/3 on 2026-07-12. Corrected.demo-screen.mdnarrated an unmerged PR Task 1.10:llm_goldensnapshot suite — pinning the whole transcript, not a property of it #11 that merged long ago.4 · README over-claims
Per your instruction, the "the reviewer is a second model session" sentence was not added.
5 · Dead scaffolding
VisionEngineandPlatformTelemetry— interfaces, fakes, providers and tests — had zero production consumers and go.lib/engines/providers.dartalso still called the app a "skeleton" and streamed "FieldOps Copilot skeleton is running on the fake engine"; it now explains what its two remaining providers are for and why the device bindings live elsewhere. Host suite: 1179 → 1177.6 · Platform and config
DEVELOPMENT_TEAMblanked at all three sites inproject.pbxproj.field_ops_copilot→ Field Ops Copilot, matching iOS.TODOs inbuild.gradle.ktsreplaced by the actual reasoning.pubspec.yamlloses the generated boilerplate and the dead fonts block.analysis_options.yamlwas the stock template with every rule commented out. It now enables the analyzer's three strict modes plus a chosen rule set, each with its reason. The strict modes pass with zero findings — no implicitdynamicleaking throughlib/despite two FFI boundaries and aSendPort. Two rules are off with a stated reason;test/andintegration_test/carry their own options file relaxing three production-lifecycle rules that misfire on a test double, whilelib/keeps all three.7 ·
docs/testing.mdrebuiltThe README sent readers there for tiers, mutation testing and the live-STT suite, and the page had none of the three. It now opens with what each tier cannot answer, states the mutation method and the three rules that make a result quotable, lists every recorded run including the one with eight survivors, and carries the CI story — the Flutter pin (a freshness gate that fires on Google's release is a gate people learn to ignore) and the codegen gate, including why
git diff --exit-codealone is not enough.Also: five dangling "see above" references from the monolith split are now real links, the three integration files the list never mentioned are added, "Tier 0/1/2" is gone everywhere, and four recursive self-correction passages are trimmed to their findings.
Added
A "Designed, not built" section in the README — eight decisions a fleet deployment forces, marked as design rather than code. Key management leads, because the honest answer is unflattering: the cipher is real and the key is a constant named
not-a-secret, which protects a stolen file and not a stolen device. The Keychain/Keystore answer already has its seam, and so does the credential-delivery equivalent one layer out.Verification
flutter analyzeclean,dart formatclean, 1177 host tests green. Generated sources regenerated so the codegen freshness gate still reports clean.