A ninth assertion that compares bytes, and the two levels it takes to land it

README.md

@@ -16,7 +16,7 @@ happened on the first fixture written.
 
 ## Read this first: the coverage this corpus does NOT have
 
-Partial coverage is easy to mistake for complete coverage. Five gaps are open.
+Partial coverage is easy to mistake for complete coverage. Six gaps are open.
 
 ### EnergyPlus 26.1.0 only
 
@@ -75,7 +75,19 @@ for. It has no members yet.
 
 A second consequence, worth stating separately: cases are curated from a sweep of the EnergyPlus
 example files, so the corpus only sees hazards that EnergyPlus's own files exhibit. Byte-order marks,
-CRLF line endings, and the other things real editors emit are not in that set.
+CRLF line endings, and the other things real editors emit are not in that set. Nine `preserve-*`
+cases were donated in `conformance-2026.9` to close part of that, and they close it for the parse
+and the write; they do not close it for the read from disk, which is this section's subject.
+
+**Assertion 9 lives on this side of the boundary too.** `preserved-text` proves that a library
+reproduces the text its read was GIVEN, not that it reproduces a file's bytes. The runners decode
+the input themselves, as above, so what the assertion holds each library to is a string in, the same
+string out. Two defects live on the other side of that line and neither is closed by it: the
+TypeScript reader refuses an IDF file carrying a byte-order mark, recorded above against
+idfkit-js#28, and Python's `save_idf` opens its destination without `newline=""`, so the standard
+library translates every line feed to the platform's own convention on the way out. On Windows a
+byte-identical write therefore becomes a byte-different file, and the first language's byte-identity
+claim is platform-dependent. Both need the reserved `checks/` directory to close.
 
 ### Weather retrieval is not covered, and `tier1` does not claim it
 
@@ -95,17 +107,21 @@ build-time warm-up and a run with the network switched off, which is where the s
 it. The `tier1` tag therefore means "the Tier 1 capabilities this corpus can express", not "every
 Tier 1 capability", and a green `--tag tier1` is not a statement about weather.
 
-### Writer output is not compared as text, and cannot be
+### The two writers' output is not compared against each other, and cannot be
 
 The naming register once said this corpus proves that both libraries render the same string for
 the same model. It does not, and it is not able to.
 
-`runners/compare.md` forbids textual comparison outright, for a good reason set out there: a
-formatting difference and a value difference are not the same finding, and a comparator that is
-textual anywhere is textual everywhere. The assertion enum has no writer kind. What assertion 3
-does instead is re-parse each library's own IDF output and compare the resulting *document* to the
-original, which catches a field that moved or a value that was lost and says nothing about the
-bytes in between.
+`runners/compare.md` forbids comparing the two libraries textually, for a good reason set out
+there: a formatting difference and a value difference are not the same finding, and a comparator
+that is textual anywhere is textual everywhere. What assertion 3 does instead is re-parse each
+library's own IDF output and compare the resulting *document* to the original, which catches a field
+that moved or a value that was lost and says nothing about the bytes in between.
+
+Assertion 9 is not an exception to this and must not be read as one. It compares a library's output
+to **that same library's own input**, never to the other library's output, and its whole content is
+that nothing was re-rendered. Rule 1 states the exception as a direction rather than as an assertion
+number precisely so the next byte comparison cannot be added by precedent.
 
 The bytes do differ, and the differences are real. Round-tripping one file through both writers
 gives 2-space indentation against 4, insertion order against sorted order, a `!-Generator idfkit`
@@ -116,6 +132,25 @@ about output belongs.
 
 So a green run is a statement about what each library understood, never about what it typed.
 
+### Three byte-hazard cases declare an assertion neither library can evaluate
+
+`preserve-empty`, `preserve-comments-only` and `preserve-duplicate-name` declare `preserved-text`
+and it is skipped in both languages, because both readers refuse their inputs and a refused read
+leaves nothing to write. `preserve-unterminated-final` is skipped on the TypeScript side alone, for
+the same reason.
+
+The first three fail for one cause between them: both libraries resolve their schema from a Version
+statement in the file, and an empty file, a file of comments and a file whose duplicate name stops
+the read have no readable model. The declared parse outcome is that refusal, which both libraries
+agree on exactly, so those cases still earn their keep on assertion 1.
+
+What they do not yet pin is the writer hazard they were donated for: that a file with nothing to
+anchor comes back whole, and that the characters of a statement the read REJECTED survive a write.
+The second needs a read that succeeds while rejecting a statement, which neither library offers
+today. The assertion is declared rather than dropped so that a library which later reads one of
+these files starts being held to the property with no corpus change, and so that the gap is visible
+as a skip in every run rather than as an absence nobody sees.
+
 ## Why `known-divergence.toml` ships populated
 
 The corpus lands with real, currently failing disagreements already recorded in

cases/encoding-latin1-high-bytes/case.toml

@@ -34,4 +34,4 @@ into one.
 tags = ["encoding"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/encoding-utf8-object-name/case.toml

@@ -33,4 +33,4 @@ reason.
 tags = ["encoding"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/extensible-empty-group/case.toml

@@ -23,4 +23,4 @@ else in the corpus catches it.
 tags = ["extensible"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/extensible-partial-group/case.toml

@@ -25,4 +25,4 @@ is a difference the oracle expectation catches.
 tags = ["extensible", "positional"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/extensible-schedule-compact-numeric-field/case.toml

@@ -23,4 +23,4 @@ reported difference rather than being rounded away.
 tags = ["extensible", "numeric", "types"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/extensible-single-vs-multiple/case.toml

@@ -21,4 +21,4 @@ multi-field group can tell those apart.
 tags = ["extensible"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/extensible-wrapper-key/case.toml

@@ -26,4 +26,4 @@ numeric coercion.
 tags = ["extensible"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/naming-blank-vs-absent/case.toml

@@ -26,4 +26,4 @@ raised it, because no shipped example leaves an optional Name blank.
 tags = ["naming", "positional"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/naming-synthetic-collision/case.toml

@@ -29,4 +29,4 @@ present only to satisfy the oracle's pre-conversion validation, so this case pin
 tags = ["naming"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/naming-synthetic-key/case.toml

@@ -22,4 +22,4 @@ present only to satisfy the oracle's pre-conversion validation, so this case pin
 tags = ["naming"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-autocalculate-casing/case.toml

@@ -20,4 +20,4 @@ schema enum names, which is not always the word the file used.
 tags = ["numeric"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-autosize-casing/case.toml

@@ -15,4 +15,4 @@ fails schema validation everywhere at once rather than in one place.
 tags = ["numeric"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-integer-field/case.toml

@@ -17,4 +17,4 @@ would turn every integer field in the corpus red without a single library changi
 tags = ["numeric", "types"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-scientific-notation/case.toml

@@ -20,4 +20,4 @@ library and only show up against an external expectation.
 tags = ["numeric"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-sentinel-schema-remap/case.toml

@@ -29,4 +29,4 @@ and that pointer is the finding.
 tags = ["numeric", "types"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/numeric-zero-against-blank/case.toml

@@ -21,4 +21,4 @@ detail of the text, not of the model.
 tags = ["numeric", "positional"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/positional-extensible-group-trailing-unset/case.toml

@@ -32,4 +32,4 @@ in the corpus rather than one standing in for the rest.
 tags = ["positional", "extensible"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/positional-trailing-unset-before-extensible/case.toml

@@ -24,4 +24,4 @@ that second failure, since it is a field-order bug and field order is compared o
 tags = ["positional", "extensible"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/positional-trailing-unset-no-extensible/case.toml

@@ -32,4 +32,4 @@ positional-extensible-group-trailing-unset.
 tags = ["positional"]
 energyplus_version = "26.1.0"
 truth = "oracle"
-assertions = ["parse-outcome", "epjson", "round-trip"]
+assertions = ["parse-outcome", "epjson", "round-trip", "preserved-text"]

cases/preserve-comment-hazards/case.toml

@@ -0,0 +1,16 @@
+title = "A comma and a semicolon inside a trailing comment, neither of which is a delimiter"
+why = """
+A comment trailing a value carries both delimiter characters. Neither is a delimiter, because a
+comment runs to the end of its line and there are no string literals in this grammar.
+
+A reader that took either for a delimiter would shift every field after it by one position, so this
+is a positional hazard reported as a comment one. For a preserving writer it is a second hazard on
+the same characters: the comment lives in the gap between two statements, belongs to no object, and
+is reproduced exactly or the file changes in a place no object was edited.
+
+Donated from the syntax fixtures of the second language.
+"""
+tags = ["positional"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]

cases/preserve-comment-hazards/input.idf

@@ -0,0 +1,6 @@
+Version, 26.1;
+
+Zone,
+  Zone One,     !- Name, which is followed here by a comma; and a semicolon
+  0.0,          !- Direction of Relative North
+  0.0;          !- X Origin

cases/preserve-comments-only/case.toml

@@ -0,0 +1,20 @@
+title = "A file of comments and nothing else, which neither reader will open"
+why = """
+Comments, no statements, and therefore no Version.
+
+For a preserving writer the property is that the output is the input: there is nothing to anchor,
+so every character is text that belongs to no object. Neither library can be asked, because both
+resolve their schema from a Version statement in the file and both refuse a file that has none. The
+declared outcome is that refusal, which the two agree on exactly.
+
+The preserved-text assertion is declared and is skipped in both languages today. It is declared
+rather than dropped so that a library which later reads a versionless file starts being held to the
+property without a corpus change, and so that the skip is visible in every run rather than being an
+absence nobody sees. The gap is recorded in the README.
+
+Donated from the syntax fixtures of the second language.
+"""
+tags = ["malformed"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]

cases/preserve-comments-only/input.idf

@@ -0,0 +1,6 @@
+! A model somebody started and never wrote a statement into.
+!-Generator IDFEditor 1.51
+!-Option SortedOrder
+!
+! There is no Version statement here either, so nothing below can be
+! resolved against a schema. Classification must still cover every byte.

cases/preserve-duplicate-name/case.toml

@@ -0,0 +1,19 @@
+title = "Two objects of one type declaring the same name, which both readers refuse"
+why = """
+Two Zone objects share a name, which the schema forbids and real files contain.
+
+Both libraries refuse the file, and they agree, which is worth pinning on its own: a duplicate name
+is the one malformed shape where the two have historically been asked to differ, and this case
+records that they do not.
+
+The writer hazard this fixture was donated for is the other half, and it is not reachable here: it
+is what happens to the characters of a statement the read REJECTED while the read still succeeded,
+which is a state neither reader offers, since both stop at the duplicate instead. The preserved-text
+assertion is declared and skipped, and the gap is recorded in the README with the other two.
+
+Donated from the syntax fixtures of the second language.
+"""
+tags = ["naming"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]

cases/preserve-duplicate-name/input.idf

@@ -0,0 +1,9 @@
+Version, 26.1;
+
+Zone,
+  Zone One,                !- Name
+  0.0;                     !- Direction of Relative North
+
+Zone,
+  Zone One,                !- Name
+  90.0;                    !- Direction of Relative North

cases/preserve-edit-no-op/case.toml

@@ -0,0 +1,19 @@
+title = "A field written the value it already holds leaves the object as it was"
+why = """
+The same input, and an operation that writes the Building's Terrain the value the file already
+gives it. Nothing changed, so nothing may be reformatted, and the operations block excludes no
+extent: the write is compared against the input in full.
+
+This case exists because it caught a real defect. The first language cleared its retained source
+text on every field write, without comparing, so this operation reformatted the Building and turned
+its `0.0` into `0`. It was found by running the library rather than by reading it, and it is
+written as a corpus case rather than as that library's unit test for exactly that reason: the
+second language got this right for free through a comparing accessor, and a claim that only one
+library's own tests hold is a claim the other can drift away from.
+
+FR-004 and the fourth acceptance scenario of the one-object-diff story.
+"""
+tags = ["numeric"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]

cases/preserve-edit-no-op/input.idf

@@ -0,0 +1,24 @@
+Version,
+  26.1;
+
+Building,
+  Scientific Notation,
+  0.0,                     !- North Axis {deg}
+  City,                    !- Terrain
+  0.04,                    !- Loads Convergence Tolerance Value {W}
+  0.4,                     !- Temperature Convergence Tolerance Value {deltaC}
+  FullExterior,            !- Solar Distribution
+  25,                      !- Maximum Number of Warmup Days
+  6;                       !- Minimum Number of Warmup Days
+
+GlobalGeometryRules,
+  UpperLeftCorner,         !- Starting Vertex Position
+  Counterclockwise,        !- Vertex Entry Direction
+  Relative;                !- Coordinate System
+
+Curve:Linear,
+  Exponent Forms,          !- Name
+  1.0E-5,                  !- Coefficient1 Constant
+  -2.5e-3,                 !- Coefficient2 x
+  -1.5E+2,                 !- Minimum Value of x
+  1.5E2;                   !- Maximum Value of x

cases/preserve-edit-one-field/case.toml

@@ -0,0 +1,18 @@
+title = "One field changes and exactly one object looks changed"
+why = """
+The input is the one from numeric-scientific-notation, unaltered. The operation writes the
+Building's Terrain, and the assertion is about every byte that is not the Building's.
+
+This is SC-002 as a case rather than as a hope. The Curve:Linear object below carries `1.0E-5`,
+`-2.5e-3`, `-1.5E+2` and `1.5E2`, four notations no writer reproduces from a parsed number, so a
+writer that reformats the whole document to satisfy one edit fails here loudly and immediately.
+The comments, the blank lines and the trailing newline are outside every object and are compared
+too.
+
+An existing input rather than a new one, deliberately: the property is about the file the operation
+did NOT touch, so the more ordinary the file the better the evidence.
+"""
+tags = ["numeric"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]

cases/preserve-edit-one-field/input.idf

@@ -0,0 +1,24 @@
+Version,
+  26.1;
+
+Building,
+  Scientific Notation,
+  0.0,                     !- North Axis {deg}
+  City,                    !- Terrain
+  0.04,                    !- Loads Convergence Tolerance Value {W}
+  0.4,                     !- Temperature Convergence Tolerance Value {deltaC}
+  FullExterior,            !- Solar Distribution
+  25,                      !- Maximum Number of Warmup Days
+  6;                       !- Minimum Number of Warmup Days
+
+GlobalGeometryRules,
+  UpperLeftCorner,         !- Starting Vertex Position
+  Counterclockwise,        !- Vertex Entry Direction
+  Relative;                !- Coordinate System
+
+Curve:Linear,
+  Exponent Forms,          !- Name
+  1.0E-5,                  !- Coefficient1 Constant
+  -2.5e-3,                 !- Coefficient2 x
+  -1.5E+2,                 !- Minimum Value of x
+  1.5E2;                   !- Maximum Value of x

cases/preserve-edit-remove-and-add/case.toml

@@ -0,0 +1,21 @@
+title = "An object removed and another added leave the text between them alone"
+why = """
+The same input as preserve-edit-rename. One Construction is removed and one Material:NoMass is
+added.
+
+Two properties in one case, and they fail in opposite directions. A removal must take the object's
+own extent and nothing else: the blank line and the comment around it belong to no object and must
+survive, and a writer that removes a line too many or too few is reproducing a file the author did
+not write. An addition must go to the end, formatted, and must not run onto the last line of a file
+whose final statement is where the author left it.
+
+The removal half is also where the object notation's terms were found to be wrong in the first
+language: its preserving write asked the surviving objects whether they were unmodified, and a
+removed object is not among them to ask, so the original text came back with the removed object
+still in it. That is fixed at the library, and this case is the text format's half of the same
+question.
+"""
+tags = ["references"]
+energyplus_version = "26.1.0"
+truth = "convention"
+assertions = ["parse-outcome", "preserved-text"]