Tests: internal/recall

internal/recall · 48 tests

Each test is named for the property it holds, and runs against a real deployment: there is no mock of the database and no arm that skips when it is absent. The sentence is the test's name read back; the name is what go test -run takes.

controls_test.go

• Recall controls are independent across concurrent clients — TestRecallControlsAreIndependentAcrossConcurrentClients
• Recall character accounting counts unicode roles and paths — TestRecallCharacterAccountingCountsUnicodeRolesAndPaths

limits_test.go

• Oversized questions are refused before accessing a store — TestOversizedQuestionsAreRefusedBeforeAccessingAStore
• An overproducing store cannot start unbounded traversal — TestAnOverproducingStoreCannotStartUnboundedTraversal

recall_test.go

• A questions windows are the shape entity names are stored in — TestAQuestionsWindowsAreTheShapeEntityNamesAreStoredIn
• An apostrophe or hyphen inside a word survives — TestAnApostropheOrHyphenInsideAWordSurvives
• A question anchors to an entity and the facts come back cited — TestAQuestionAnchorsToAnEntityAndTheFactsComeBackCited. ── #3's claim ──────────────────────────────────────────────────────────────────────────────── Observe, form, recall — and the fact comes back with its verbatim quote and byte span.
• An entity is an anchor from either end of a fact — TestAnEntityIsAnAnchorFromEitherEndOfAFact. A fact is reachable through EITHER end.
• A recall sees only the projects it was authorised for — TestARecallSeesOnlyTheProjectsItWasAuthorisedFor. ── The soft boundary ───────────────────────────────────────────────────────────────────────── A project is a permission, so the authorised set is an argument on every read.
• An empty authorised set returns nothing — TestAnEmptyAuthorisedSetReturnsNothing. An empty authorised set returns nothing.
• A question that anchors to nothing is an empty bundle — TestAQuestionThatAnchorsToNothingIsAnEmptyBundle. A question naming nothing we know about is an empty bundle, not an error.
• A truncated bundle says so — TestATruncatedBundleSaysSo. A truncated bundle says so.
• A nonsense bundle cap falls back rather than returning nothing — TestANonsenseBundleCapFallsBackRatherThanReturningNothing. A bundle cap of zero or less falls back rather than returning nothing.
• Facts from one message about one subject are marked not independent — TestFactsFromOneMessageAboutOneSubjectAreMarkedNotIndependent
• Facts from different messages of one turn are independent — TestFactsFromDifferentMessagesOfOneTurnAreIndependent. Two messages in one turn are two speakers.
• Facts about different subjects are independent — TestFactsAboutDifferentSubjectsAreIndependent. Different subjects in one sentence are different claims.
• Co derivation compares subjects the way entities resolve — TestCoDerivationComparesSubjectsTheWayEntitiesResolve. The subject is compared the way entities are resolved, so casing does not make one person two.
• Co derivation names only what the caller can see — TestCoDerivationNamesOnlyWhatTheCallerCanSee. Only what is in the bundle is named.
• A fact with no subject is not co derived with everything — TestAFactWithNoSubjectIsNotCoDerivedWithEverything. A fact whose subject or evidence is gone — an erased entity — is independent of nothing rather than co-derived with everything, which is what an empty key would make it.
• A question naming nothing known says so rather than looking silent — TestAQuestionNamingNothingKnownSaysSoRatherThanLookingSilent. A question that names nothing this memory holds says so, rather than returning an empty bundle that looks like having nothing to say.
• A known subject with no facts is not the same as an unknown one — TestAKnownSubjectWithNoFactsIsNotTheSameAsAnUnknownOne. A question that names something known and finds no facts is the OTHER empty answer, and must not claim the thing is unknown.
• The bundle says where its facts came from — TestTheBundleSaysWhereItsFactsCameFrom. Where the facts came from is visible, so a question that named several things and drew everything from one is recognisable without anybody choosing a threshold.
• The counts describe what the caller got — TestTheCountsDescribeWhatTheCallerGot. The counts describe what the caller received.
• The bundle is cut by what the caller can afford — TestTheBundleIsCutByWhatTheCallerCanAfford. The cut is by characters, not rows, because a row is not a size: one fact carries a two-word quote and another carries a paragraph, so a caller converting rows to anything has to assume the worst every time to avoid overrunning once.
• A budget too small for one fact returns an explicitly truncated empty bundle — TestABudgetTooSmallForOneFactReturnsAnExplicitlyTruncatedEmptyBundle. A too-small context allowance returns no facts and explicitly reports the budget cut.
• A generous budget returns everything and says so — TestAGenerousBudgetReturnsEverythingAndSaysSo. A generous budget returns everything and does not claim truncation.
• The row bound guards the query — TestTheRowBoundGuardsTheQuery. The row bound guards the query rather than the answer: an enormous budget must not become an unbounded read.
• The words around a quote are charged to the budget — TestTheWordsAroundAQuoteAreChargedToTheBudget. The context is on the wire, so the caller pays for it.
• A bundle says when a citation lost its context — TestABundleSaysWhenACitationLostItsContext. A bare citation is announced once for the bundle.
• An ordinary recall names no instant — TestAnOrdinaryRecallNamesNoInstant. The ordinary recall asks about now, and "now" is the zero value rather than a timestamp this process computed — the store turns that into the open interval, which is the only version that does not depend on two clocks agreeing.
• A historical recall passes both instants through — TestAHistoricalRecallPassesBothInstantsThrough. And a historical recall is the same read with two instants on it: same anchoring, same scopes, same sources, same budget.
• An ordinary recall asks for one hop — TestAnOrdinaryRecallAsksForOneHop. The ordinary recall asks for one hop.
• A depth beyond the bound is clamped rather than refused — TestADepthBeyondTheBoundIsClampedRatherThanRefused. A depth beyond what the traversal does is clamped rather than refused: a caller asking for more memory gets the deepest answer the system gives, which is the useful outcome.

surfaces_test.go

• An unanchored question names the surfaces that did not run — TestAnUnanchoredQuestionNamesTheSurfacesThatDidNotRun. With no exact anchor and no semantic surfaces, a caller who selected surfaces by name is told which did not run, rather than getting an empty answer that looks like ignorance.
• A semantic anchor starts a walk when no name matched — TestASemanticAnchorStartsAWalkWhenNoNameMatched. A question that names nothing stored by name is anchored by meaning, and the anchor says so.
• A thematic question with no anchor is answered from reports with sources — TestAThematicQuestionWithNoAnchorIsAnsweredFromReportsWithSources. A thematic question with no anchor is answered from reports, each carrying its sources, and the reach still says nothing was anchored because that is still true.
• Passages fill the remaining budget and are never facts — TestPassagesFillTheRemainingBudgetAndAreNeverFacts. Passages fill what the budget has left and are never facts: a tool message's words come back as a passage with its role, and the facts list is untouched by them.
• A refusing surface is degraded not fatal — TestARefusingSurfaceIsDegradedNotFatal. A surface that refuses is named as degraded and the rest of the bundle still answers.
• The composed surfaces share one budget — TestTheComposedSurfacesShareOneBudget. The budget is one budget: reports and passages are cut by the characters the facts left.
• Surface controls are refused when not in the set — TestSurfaceControlsAreRefusedWhenNotInTheSet. Surface names are a closed set: an unknown or repeated name is refused, and the exact path is not touched when only facts are selected and a name matched.
• Semantic anchors are capped and stop at the cap — TestSemanticAnchorsAreCappedAndStopAtTheCap. A meaning match proposes at most a fixed number of places to start, however many scopes are asked: the first scope that fills the cap ends the search, so a caller with many projects does not pay one embedding query per project for anchors it will not walk.
• A report larger than the remaining budget is cut whole and reported — TestAReportLargerThanTheRemainingBudgetIsCutWholeAndReported. A report that does not fit what the budget has left is cut whole and the cut is reported: half a report is not a smaller report, it is a different one.
• A refusing passage surface is degraded and reports still answer — TestARefusingPassageSurfaceIsDegradedAndReportsStillAnswer. The passage surface refusing is named as degraded like any other, and the reports still answer.
• No scopes answer nothing and a repeated word is one term — TestNoScopesAnswerNothingAndARepeatedWordIsOneTerm. Nothing to search is an empty answer, not an error, and a repeated word is one term: a term is a thing to look up, and looking the same thing up twice is the same lookup.
• A subject filter reaches every surface — TestASubjectFilterReachesEverySurface. Passages are searched for that person only; reports, which span people and name nobody as their source, are withheld and the bundle says so; and a name the semantic surface proposed that this person has no fact about is not returned, because it is somebody else's memory.
• Deselecting facts returns no facts even when a name matched — TestDeselectingFactsReturnsNoFactsEvenWhenANameMatched. Deselecting facts deselects them even when the question named something: no walk, no anchors, no facts — and the name it matched is still counted, so the reach does not claim it named nothing.
• A semantic anchor that answers is counted as anchored — TestASemanticAnchorThatAnswersIsCountedAsAnchored. A bundle answered through a semantic anchor counts that anchor, so it does not also report that the question named nothing this memory holds; under a subject filter the anchor stays, because that person has a fact about it.
• A fact is never co derived with itself — TestAFactIsNeverCoDerivedWithItself. The same fact twice is one fact, never co-derived with itself.