Restrict --tfm all to package discovery

[richlander]

Part of #9833 under #9240. This corrects #9904 and supersedes the document-only rule from #3885.

Summary

--tfm all is package discovery, in the manner of NuGet Package Explorer. It lists archive content across every framework by path and never touches the IL reader. Introspection across several frameworks has no meaningful answer, so --tfm all is accepted only by package. Every request that needs one selected framework, or reads an assembly or PDB, fails before acquisition.

Design basis:

• Normative owner: a new focused document, All-frameworks package discovery.
• Inventory: it consumes compile selection's existing available-slice inventory, and that contract is unchanged.
• Cross-references: the other owners only point to it.

Changes

• package-tfm-all-discovery.md (new owner)
  • Exact claim and admission table:
    • The default request returns a discovery projection. Files, Dependencies (every group) and archive- or metadata-only sections are admitted.
    • Package Info is admitted without its selected-framework, ecosystem and binary_signals fields.
    • --library, Dependency Hierarchy, assembly- or PDB-reading sections, and categories containing them fail.
    • package query fails.
    • Every non-package command with a package target fails and names package <id> --tfm all.
    • Every non-package target (--platform, local file, project) fails.
    • Admitted sections skip the package-wide binary-signal scan.
  • Discovery projection:
    • It is a Hierarchy of rows with framework, role and path cells, under framework context nodes ordered by TfmResolver priority.
    • Every candidate is a row. Explicit _._ markers are rows. Tool slices come from Package Info tool measurements, and a tool slice without a Library is also a row.
    • Tree is the native presentation. JSON, envelope, TSV, JSONL, table, Count and --rows observe the same rows.
    • Only rows in their framework's selected compile projection, and tool Library rows, are followed with --tfm <framework> --library <path>. The rest are discovery-only.
  • Adoption:
    • It lists observed current behavior and the gates that change.
    • It names the documents whose --tfm all descriptions and gates retire under Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833: integrations, inspection space, assembly inspection query, CLI row selection, CLI architecture, find search service, untrusted data threat model, assembly audit, artifact acquisition, command transition model, CLI reference and primary subject views.
• output-shapes.md: the Declare library --tfm all as one Library aggregate Table #9904 aggregate Table contract is replaced by a cross-reference.
• section-cardinality.md: --tfm all has no Library row unit.
• package-file-inventory.md: "--tfm all adds no predicate", and admission is deferred to the new owner.
• package-library-scope.md: the children document is scoped to requests without --tfm all, and the --tfm all Library routes are retired rather than repopulated.
• package-info-tool-measurements.md: the CLI all gate's selected-TFM assertions change with the adoption.
• artifact-acquisition-and-workspaces.md: its two all-framework gate cases become rejection gates.

Not in this PR:

• the CLI adoption (Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833);
• the NoCompileAssets Library population decision (Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833 and Package library scope).

Validation

• npx markdownlint-cli on all seven changed files: clean.
• Current behaviors were observed with a NativeAOT build of main. They include EntityFramework@6.4.4, System.Linq@4.3.0, Microsoft.CSharp@4.7.0, dotnet-ef and tool packages.
• Every doc, skill and the README were swept for all spellings of --tfm all. The remaining "all-framework" mentions describe the router's installed-framework platform search, which this change does not touch.
• 11 adversarial review rounds. The maintainer approved continuing past the six-round checkpoint. Blocking findings per round went 2, 3, 2, 2, 1, 1, 1, 2, 1, 1, and round 11 was clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_018t8FB3ipLLs41ECJvLVNfu

[richlander]

Round 1 reconciliation (candidate 0c5188e91)

The adversarial review (one seat) found 2 blocking findings. Both are accepted and fixed in 293acad56.

Finding	Disposition
F1. "Never acquires assembly metadata" and "valid only on package discovery" conflict with package sections that still read every DLL under --tfm all. Examples: -S @SourceLink (8 libraries, PDB rows), -S Signals, and Package Info's selected-TFM facts. type --package P --tfm all fails after acquisition instead. None of these was in the adoption list.	Accepted. The rule now states its admission criterion. A request that consumes a selected compile projection or reads an assembly fails before acquisition and names package <id> --tfm all. That covers library/type/member package targets, package --library, and assembly-reading package sections (SourceLink, Signals, Package Info's selected-framework facts). Sections that read only archive entries or package metadata are admitted. The #9833 adoption list names every route.
F2. The all-framework population was assigned to package-library-scope.md, which excludes Package selection and the native Package Tree. package-file-inventory.md's own scope is Files.	Accepted. Compile selection owns target requests and already issues the binary-free available-slice inventory, so it now owns --tfm all: --tfm all is not a target request and selects no projection. The default hierarchy shows every available slice, in the inventory's deterministic order, with its ref/lib candidates by path and its explicit empty-group marker. package-file-inventory.md, output-shapes.md, section-cardinality.md and package-library-scope.md now only cross-reference it.

Non-blocking observations:

• Duplicate restatement: output-shapes.md now carries a one-sentence cross-reference instead of repeating the rule.
• Stale current-mechanics docs: integrations.md, inspection-space.md, assembly-inspection-query.md, and the cli-row-selection.md exclusion lists go stale once Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833 removes the code.
• Code strings: the CLI help at LibraryAddressCommandDefinitions.cs:464 and the error strings at LibraryCommand.cs:3780-3817 also change in that adoption.

Next: round 2 on 293acad56.

---

Generated by Claude Code

[richlander]

Round 2 reconciliation (candidate 293acad56)

The adversarial review (one seat) found 3 blocking findings. All three are accepted and fixed in 64ae94b46.

Finding	Disposition
B1. Compile selection's Composition section disclaims CLI behavior and rendering. It also has no inventory-only request and doesn't cover tool packages. So it could not own command admission, error text or hierarchy rendering, which repeats F2 at a different owner.	Accepted. A new focused owner, All-frameworks package discovery, owns --tfm all. It consumes compile selection's existing available-slice inventory without changing that contract. The compile-selection edit is reverted, and the other four docs now only cross-reference the new owner.
B2. The "reads an assembly" test is the wrong criterion. -S Dependencies --tfm all returns 0 rows because it filters by the literal all. -S "Dependency Hierarchy" --tfm all throws an internal exception. Neither section was classified.	Accepted. Admission now fails anything that "needs one selected framework, or reads an assembly or a PDB". The owner's admission table covers: Dependencies: every dependency group. Dependency Hierarchy: fails before acquisition. Package Info: drops its selected-framework fields. Manifest-only sections: unchanged. A category containing a failing member: fails rather than dropping members.
B3. find, extensions, diff and depends accept --tfm all today and weren't in the covered list or the adoption list.	Accepted. Every command with a package target other than package fails before acquisition and names package <id> --tfm all. The adoption section lists the observed find, extensions and diff behavior.

Non-blocking observations:

• Package Info: resolved as "the section without its selected-framework fields".
• Validation site: the adoption now says Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833 adds one shared pre-acquisition check. Today library and type recognise all at separate sites.
• Tool packages: the hierarchy shows every tool slice by path.
• Stale current-mechanics docs: named in the owner's adoption section for Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833 to retire.
• package-library-scope.md paragraph: reordered so "Observed results" refers to the archive search again.

Next: round 3 on 64ae94b46.

---

Generated by Claude Code

[richlander]

Round 3 reconciliation (candidate 64ae94b46)

The adversarial review (one seat) found 2 blocking findings. Both are accepted and fixed in 1e2ec9ea9.

Finding	Disposition
R3-B1. Package Info still admitted fields that read assemblies for a selected framework: Ecosystem Dependencies and Ecosystem Dependency Status (the rollup of the Ecosystem Dependencies section, which fails), and Selected-TFM Status.	Accepted. Package Info now omits every field that needs a selected framework or reads an assembly, and lists them: Selected TFM, Selected-TFM Status, the selected-TFM folder/library count/size measurements, Ecosystem Dependencies and Ecosystem Dependency Status.
R3-B2. The --tfm all default Tree overlapped the Package children document owned by package-library-scope.md. Its rows promise package P --library <path> navigation, and the default framework rejects other frameworks' paths.	Accepted. The --tfm all default Tree is now a discovery projection the new owner defines, not the children document. It is grouped by framework in descending priority, then by name. It shows ref/lib candidates by path and explicit empty markers. Tool packages group their tool slices per Package Info tool measurements. Every row carries its framework, so following a row uses --tfm <framework> --library <path>. package-library-scope.md records that boundary in one sentence, and its children-document population is unchanged.

Non-blocking observations:

• Taken: section rows. A row for "sections that read only archive entries or package metadata" replaces the "manifest-only" label.
• Taken: tool-slice owner. It is now linked, and the projection owns its own order.
• Taken: adoption bullet corrected. Package-mode depends already fails before acquisition. Type-mode depends, member and workspace accept --tfm all, and package query returns zero matches.
• Taken: package query classified. It fails before acquisition because it evaluates assemblies. workspace is added to the failing commands.
• Taken: stale docs. cli-architecture.md and find-search-service.md are added to the adoption's stale-docs list.
• Not taken: docs/overview.md authority entry. That file is release-managed, so its owner entry rides release preparation. The recent package-domain owners also have none.
• Not taken: --library error text. The row's "names the section or flag" stands, and the frame wording was mine.

Next: round 4 on 1e2ec9ea9.

---

Generated by Claude Code

[richlander]

Round 4 reconciliation (candidate 1e2ec9ea9)

The adversarial review (one seat) found 2 blocking findings. Both are accepted and fixed in 586a27d88.

Finding	Disposition
R4-B1. The discovery projection owned only the default Tree. Under --tfm all, JSON, the envelope, row formats, Count and --rows were still bound to the selected-framework children document. On main they report target_framework: netstandard2.1, Count 2, and netstandard2.1-only TSV.	Accepted. The admission row is now "default request, in every format and terminal". The projection is a Hierarchy under Section shapes: Its rows are candidate assets with framework, role and path cells. They sit under framework context nodes, ordered by TfmResolver priority (highest first), then by name. A slice with an explicit empty group shows its marker and contributes no rows. Tool slices come from Package Info tool measurements. An empty inventory returns the typed no-Library state with zero rows. Tree is the native presentation. --tree, JSON, the envelope, TSV, JSONL and table output lower the same rows, Count counts them and --rows windows them. package-library-scope.md's children-document sentence is now scoped to requests without --tfm all. The adoption names the QuerySpace producer declaration, cli-reference.md and primary-subject-views.md.
R4-B2. Admitted sections still run the package-wide binary-signal scan (PackageInspector.InspectAsync, scanBinarySignals), which opens every DLL and PDB. Package Info JSON emits the result as binary_signals.	Accepted. Admitted sections now produce their ordinary results without the binary-signal scan, and Package Info excludes binary_signals and every member derived from it. Adoption records the scan site, and #9833 skips the scan under --tfm all.

Non-blocking observations, taken:

• Rank: named as TfmResolver priority.
• Empty inventory: stated.
• NoCompileAssets non-claim: clarified. Under --tfm all, such a package shows the empty state and is not searched.

Next: round 5 on 586a27d88. Round 6 is the checkpoint.

---

Generated by Claude Code

[richlander]

Round 5 reconciliation (candidate 586a27d88)

The adversarial review (one seat) found 1 blocking finding. It is accepted and fixed in 40e1e8282. Everything from rounds 1-4 is confirmed resolved.

Finding

Disposition

R5-B1. "A slice with an explicit empty group … contributes no rows" had two effects. It dropped real candidates from a slice that also has a marker, which compile selection allows. It also left marker-only frameworks in the Tree but absent from flat formats and Count. System.Linq@4.3.0 has 11 marker-only slices and 4 real ones: the Tree would show 15 frameworks, while TSV and Count would reflect 4.

Accepted. Every candidate is a row, whatever markers its slice also has. Each explicit empty-group marker is itself a row with role ref and archive path ref/<tfm>/_._, so every framework appears in every lowering and in Count.

Non-blocking observations:

• Taken: empty state. It now applies to an empty compile and tool inventory.
• Taken: producer placement. The producer sits beside the Package children producer with its own QuerySpace declaration. It has no Browser/Wasm consumer, because Inspect Web has no --tfm all gesture.
• Taken: --library navigation. In package-library-scope.md, "Without --tfm all," now scopes the navigation sentence.
• Left: binary-scan adoption bullet. It is slightly broad, since the lightweight Target Frameworks path and cache hits don't rescan. The contract is unaffected.
• Left: Audit Findings not named. The "reads an assembly or a PDB" criterion already classifies it.

Next: round 6 on 40e1e8282, which is the six-round checkpoint.

---

Generated by Claude Code

[richlander]

Round 6 reconciliation (candidate 40e1e8282): six-round checkpoint

The adversarial review (one seat) found 1 blocking finding. Every finding from rounds 1-5 is confirmed resolved, and the reviewer reports nothing else blocking.

Finding

Disposition

R6-B1. The rule "every row is followed with --tfm <framework> --library <path>" contradicts the row population. Rows include lib candidates shadowed by a real ref asset, and _._ marker rows, and neither is a valid --library target. On Microsoft.CSharp@4.7.0, --tfm netstandard2.0 --library lib/netstandard2.0/Microsoft.CSharp.dll and --tfm net45 --library ref/net45/_._ both exit 1. The lib-shadowing half was latent since round 4, and round 5's marker rows made the conflict unavoidable.

Accepted, not yet fixed. Proposed one-sentence edit: a row is followed with --tfm <framework> --library <path> only when its path is in that framework's selected compile projection, or it is a tool row. Marker rows and shadowed or suppressed candidates are listed for discovery only. The R5-B1 population is kept, and compile selection is unchanged.

Non-blocking, left:

• The marker path is rebuilt in the inventory's framework spelling, not the archive's exact casing.
• lib/<tfm>/_._-only frameworks are not compile slices, so the projection omits them (-S Files still shows them).

The PR has reached the repository's six-round checkpoint, so I'm stopping here for a decision on further rounds.

---

Generated by Claude Code

[richlander]

The maintainer approved continuing past the six-round checkpoint. 9bb129528 fixes R6-B1:

• Navigable rows: only a row whose path is in its framework's selected compile projection, or a tool row, is followed with --tfm <framework> --library <path>.
• Discovery-only rows: marker rows, and candidates that a real ref asset shadows or a co-located marker suppresses, are listed for discovery only.
• Marker spelling: marker paths are stated in the inventory's framework spelling.

Round 7 review is starting on 9bb129528.

---

Generated by Claude Code

[richlander]

Round 7 reconciliation (candidate 9bb129528)

The adversarial review (one seat) found 1 blocking finding. It is accepted and fixed in 44c1fdd33. R6-B1 is confirmed resolved, and the reviewer found nothing else blocking.

Finding	Disposition
B1. package-info-tool-measurements.md has a required gate, PackageCommand_DeclaredToolUsesAggregateToolMeasurementsColdAndWarm, that runs -S "Package Info" --tfm all and asserts the selected-TFM fields. The new owner omits those fields under --tfm all, without a cross-reference or an adoption entry.	Accepted. The tool-measurements owner now records, below its gate table, that the CLI all gate's selected-TFM assertions change with the All-frameworks discovery adoption under #9833. Its download and cold/warm assertions are unchanged. The new owner's Adoption section lists current Package Info behavior (selected fields and Ecosystem Dependencies are kept) and that gate.

Non-blocking, taken: a tool slice without a Library is now a row with role tool and its slice directory path, so it stays visible in every format. That mirrors the R5-B1 marker fix.

Next: round 8 on 44c1fdd33.

---

Generated by Claude Code

[richlander]

Round 8 reconciliation (candidate 44c1fdd33)

The adversarial review (one seat) found 2 blocking findings. Both are accepted and fixed in 7cfd2108a. Everything from rounds 1-7 is confirmed resolved, and the whole-document check found nothing else blocking.

Finding	Disposition
B1. Non-package targets (--platform, a local file) accept --tfm all, under three gated tests: BodyShapesSectionTests.TfmAll_StaysOutsideTheDeclaration, LibraryCommand_TfmAll_PlatformRouteRetainsSingleShapeOutput and LibraryCommand_TfmAll_LocalFileRetainsSingleShapeOutput. cli-row-selection.md also records a gate row for them. The exact claim says only package accepts --tfm all, but the table didn't classify these routes.	Accepted, following the maintainer's direction that --tfm all should only be on package. A new admission row says any command with a non-package target fails before acquisition and states that --tfm all applies only to package. Adoption records the three gates and the row-selection gate row, which the #9833 adoption replaces with rejection gates.
B2. untrusted-data-threat-model.md and assembly-audit.md name LibraryCommand_TfmAll_PreservesHealthyIdentifierAuditResults as a per-framework audit gate, and Adoption didn't list them.	Accepted. Both are now in Adoption's retirement list.

Next: round 9 on 7cfd2108a.

---

Generated by Claude Code

[richlander]

Round 9 reconciliation (candidate 7cfd2108a)

The adversarial review (one seat) found 1 blocking finding. It is accepted and fixed in 9c1c8fed1. Everything from rounds 1-8 is confirmed resolved.

Finding	Disposition
B1. artifact-acquisition-and-workspaces.md gates package --library --tfm all as a supported route. One gate is the all-framework case of PackageCommand_ExplicitTfmPreservesSelectionAndUsesCompatibleArtifactRoles; the other is the all-frameworks case of PackageCommand_LocalInspectionSelectionPreservesSupportedShapes. Adoption didn't name that doc.	Accepted. That owner now records beside its gate that both cases become rejection gates under the new owner. The discovery owner's Adoption lists the current behavior and the gate owner. It also adds the command-transition-model.md all-TFM row, which was non-blocking in the review.

Sweep: to stop finding these one per round, I searched every doc, skill and the README for --tfm all, all-framework, all-TFM, TfmAll, AllTfm and 'all'.

• Every remaining owner is now referenced by the new doc.
• The leftover "all-framework" mentions in cli-platform-type-routing.md and member-inspection-planning-and-metadata-projection.md are a different thing: the router's installed-framework platform search. This change doesn't affect them.

Next: round 10 on 9c1c8fed1.

---

Generated by Claude Code

[richlander]

Round 10 reconciliation (candidate 9c1c8fed1)

The adversarial review (one seat) found 1 low-severity blocking finding. It is accepted and fixed in 2564cd8ac. The reviewer re-ran the doc/skill sweep and confirmed that every --tfm all owner is covered, and that the platform "all-framework" mentions are out of scope.

Finding	Disposition
R10-B1. A tool-slice row without a Library was both followable (it is a "tool row") and implicitly discovery-only (its directory path is no --library target).	Accepted. Navigation now applies to tool Library rows. Tool-slice rows without a Library join markers and shadowed or suppressed candidates as discovery-only rows.

Non-blocking, noted for #9833:

• Several --tfm all tests are not named in any doc; they retire with the IL-reading paths.
• The --library error wording is governed by the owner's row ("names the section or flag").

Next: round 11 on 2564cd8ac.

---

Generated by Claude Code

[richlander]

Round 11 reconciliation (candidate 2564cd8ac)

The adversarial review (one seat) came back clean at 2564cd8ac. Every finding from rounds 1-10 is confirmed resolved, the ownership split is clean, and no target-contract text still treats library or package --library --tfm all as supported.

Non-blocking, left as they are:

• Coverage of the two row lists. Neither the followable list nor the discovery-only list covers every candidate on its own, for example nested candidates. The closing guarantee ("no row implies a --library path that its own framework's selection would reject") decides those cases.
• Tool package that also has compile slices for one framework. This would get two context nodes, which is practically nonexistent. It is noted for the Package Library population: one compile view, explicit RID implementation, no whole-package fallback #9833 adoption.
• Removed Declare library --tfm all as one Library aggregate Table #9904 text. Dropping it restores Package --library to its pre-Declare library --tfm all as one Library aggregate Table #9904 contract.

---

Generated by Claude Code