sandwich-hime/docs/SECURITY_EVIDENCE.md

Security evidence ledger

This ledger records what was actually inspected and executed. It is a maintainer-led self-assessment, not an independent audit, certification, formal verification, or guarantee that no vulnerability exists.

Assessment identity

Field	Value
Assessment date	2026-08-12
Evidence sets	Clean security self-assessment plus exact-commit Beta 1 platform, release, signing, and installation checks
Public commit	b7a84054d755e42285e50298e41e47f06a8325a5
Public tree	be9e118e38dfebed19f60403ededdadabe07d2aa
Maintainer-run environments	Windows 11/amd64 on NTFS; Ubuntu 20.04/amd64 under WSL2 on ext4; supplementary pre-beta Linux/amd64 server containers
Supported Go lanes exercised	Go 1.25.12 and Go 1.26.5
Declared minimum Go	Go 1.25
Assessor	Project maintainer with AI-assisted code review; human responsibility retained

The named Windows and WSL2 platform runs used the exact public commit and tree above. The isolated server-container matrix preceded the final candidate and is retained only as supplementary Linux evidence. Hostnames, network addresses, account names, private paths, private repository identities, and private commit mappings are intentionally absent from this public ledger. These platform observations are historical evidence, not the current support matrix. Linux/amd64 and Darwin/arm64 are now the maintained v1 release targets. This section retains historical Beta 1 evidence; the exact RC must supply new native evidence on both targets. WSL and native Windows are not v1 release blockers.

Beta 2 compiler publication addendum

This addendum records the additive language-server release without replacing the Beta 1 assessment identity above. Compiler tag v1.0.0-beta.2 is a signed annotated tag whose peeled public commit is 1082d9d61eb84e67ca4012ff9ee3898ee37ac6fd and whose public tree is 01d5702928f3d9c9fb0e3d2213530add7ff94745. The tag object is f091cd67f688ba5ee784b18f5a407a9326df7ab2.

Beta 2 changed only the development compiler. No sando/v1.0.0-beta.2 tag was created. The release preflight compared the retained runtime tag sando/v1.0.0-beta.1 with the Beta 2 commit and verified that their sando subtrees were byte-identical at tree 3035e948f77f160d399089be3ae80c88bab3fed2.

The exact public Beta 2 source passed the full race-enabled verifier and the compiler-only release preflight on executed Linux with Go 1.26.5. Fresh native Windows checkouts on NTFS passed the full race-enabled PowerShell verifier, focused process-tree/watcher/consumer tests, candidate-stamped version checks, and deterministic generation on Go 1.25.12 and Go 1.26.5. Clean isolated GOPROXY=direct and public-proxy-only installs produced features:["lsp-stdio"]; the public-proxy path also verified the retained runtime through sum.golang.org. The Windows result is retained as historical portability evidence and does not create an ongoing support promise.

The Beta 2 language server is additive development tooling. Its tested security boundary includes protocol-only stdout; bounded header and message framing; integer/string JSON-RPC identifiers; full-document in-memory overlays; UTF-16 conversion at the protocol boundary; cancellation and shutdown; workspace, nested-module, VCS, symlink, and file-count boundaries; and explicit no-write/no-network/no-Go-tool execution tests. Fuzz targets exercise bounded JSON-RPC framing and document changes. These checks do not make an untrusted workspace safe to execute: himesan dev and project commands remain trusted local-code operations, while himesan lsp --stdio performs analysis only.

Observed security self-assessment evidence

These commands were observed on clean remediated source during the dated assessment. Except where the exact-commit platform matrix below says otherwise, the table does not claim that every command was rerun on the named public baseline commit.

Property examined	Enforcement or test surface	Result observed on 2026-08-12
Root correctness	go test -count=1 ./...	Pass
Concurrent access	go test -race -count=1 ./...	Pass
Runtime concurrency	(cd sando && go test -race -count=1 ./...)	Pass
Standard static analysis	go vet ./... and runtime equivalent	Pass
Reachable known vulnerabilities	govulncheck@v1.6.0 on both modules	No vulnerabilities found on 2026-08-12
Dependency surface	go list -m -json all in both modules	Zero third-party module requirements
Statement coverage	Go cover profiles on remediated source	compiler 75.2%; devserver 80.2%; runtime 96.6%
Parser robustness smoke	Two bounded Go fuzz targets	Pass; no panic found
Deterministic generation	repeated generate/check/hash/mtime gates	Pass
Writer failures	runtime error/short-write/nil-writer tests	Pass
HTML text/attribute/RCDATA escaping	compiler/runtime adversarial cases plus the committed html/template overlap corpus	Pass for the committed corpus; documented stricter invalid-UTF-8 handling remains intentional
URL scheme handling	ordinary/trusted URL matrices plus safe, unsafe, and intentionally divergent html/template cases	Pass for the committed corpus; control rejection and the explicit tel allowlist are documented policy differences
Filesystem boundaries	symlink, nested-module, VCS, ownership, stale-output tests	Pass for tested cases; see open findings
Development proxy browser boundary	Host, Origin, Fetch Metadata, CSP, fragment and response tests	Pass for tested cases
Platform behavior	Historical exact-candidate native Windows and executed Linux matrices	Windows/Linux passed for the tested lanes; the v1 RC requires fresh Linux/amd64 and Darwin/arm64 evidence

Coverage measures statements executed by tests. It is not branch completeness and is not evidence that the executed behavior is secure.

govulncheck reports vulnerabilities known to the Go vulnerability database and reachable through its analysis. A clean result cannot detect unknown flaws, design errors, or vulnerabilities outside its model.

Historical Beta 1 compatibility matrix

These are maintainer-run, point-in-time results, not continuous CI and not an independent audit.

Environment	Go lanes	Commands and focused evidence	Result and limits
Windows 11/amd64, NTFS	1.25.12, 1.26.5	Native PowerShell verifier with race; root/runtime tests, vet, trimpath build, freshness, two generation passes, process-tree cleanup, watcher boundaries, and temporary consumer compilation	Pass. Symlink-output rejection skipped because the test account lacked symlink privilege; the read-only-directory case is POSIX-only
Ubuntu 20.04/amd64 under WSL2, native ext4 checkout	1.25.12, 1.26.5	Race-enabled verifier; root/runtime tests, vet, build, two generation passes, ten focused filesystem cases, five focused development-process/watcher cases, and license check	Pass. This is Linux execution under WSL2, not bare-metal or Linux/arm64 evidence
Linux/amd64 server containers	1.25.12, 1.26.5	Earlier pre-beta root/runtime tests, vet, builds, race, licensing, and deterministic generation in sequential isolated official Go containers	Pass on the earlier baseline only. Container resources were capped at 1 CPU and 2 GiB; this is supplementary evidence, not an exact Beta 1 lane or Linux/arm64 evidence
macOS	—	Cross-compilation only	No native Beta 1 evidence; Darwin/arm64 becomes a maintained target at the v1 RC and requires fresh evidence

The generated golden basic.sando.go was 1,399 bytes and had SHA-256 63fa75a3049a3a8a12d769d7f9b6b510dfe763baacf706775b75cef2c57a984f on every tested Windows and Linux lane. Repeated generation also preserved its timestamp. This demonstrates cross-host agreement for one compiler-owned fixture, not equivalence for every possible template.

Portability reports for unsupported targets may include the operating system, architecture, go version, exact command, and a minimal reproduction. Suspected vulnerabilities use the private route in SECURITY.md. Such reports help find gaps but do not create a support promise; maintainers remain responsible for security triage and fixes on both supported native targets.

Security-relevant design evidence

Production boundary

The production sando module contains rendering contracts and contextual write helpers. It contains no HTTP server, router, middleware, template discovery, development proxy, plugin loader, or production process manager. Both compiler and runtime modules currently have no third-party Go module requirements.

Compiler behavior

Compilation builds and formats outputs in memory before generation writes. Recursive discovery rejects or skips observed symlinks, nested modules, VCS trees, vendor trees, and detected filesystem crossings. Existing non-owned, symlink, and non-regular output files are rejected. Each changed file uses an atomic replacement primitive; the whole set is not a filesystem transaction if a later replacement fails.

generate and check do not run project code, invoke the Go toolchain, fetch dependencies, or edit module metadata. himesan dev is intentionally separate: it builds and executes trusted project code and may fetch modules under the user's normal Go configuration.

Output contexts

The compiler accepts dynamic values only in its enumerated contexts. It rejects dynamic markup construction, unquoted attributes, event-handler values, dynamic style attributes, unsupported URL lists, foreign content, meta refresh, malformed tags, and unbalanced generated components. Runtime helpers escape ordinary text/attributes, validate ordinary whole-URL schemes before writing, and escape every trusted wrapper in RCDATA.

Script and style output require opaque trusted types. Those types deliberately move responsibility to trusted application code; they are not sanitizers.

Reproduction commands

Run from a clean canonical checkout. Networked scans contact the Go module proxy and vulnerability database.

go version
git status --short
git rev-parse HEAD^{commit} HEAD^{tree}

./scripts/check-licenses.sh
HIMESAN_RACE=1 ./scripts/verify.sh

go test -count=1 -cover ./...
(cd sando && go test -count=1 -cover ./...)

go run golang.org/x/vuln/cmd/govulncheck@v1.6.0 ./...
(cd sando && go run golang.org/x/vuln/cmd/govulncheck@v1.6.0 ./...)

go test ./internal/compiler -run '^$' \
  -fuzz '^FuzzCompileNeverPanics$' -fuzztime=20s
go test ./internal/compiler -run '^$' \
  -fuzz '^FuzzGoDelimiterNeverPanics$' -fuzztime=20s

./scripts/release-check.sh --version v1.0.0-beta.1
./scripts/verify-public-install.sh --version v1.0.0-beta.1

Those historical Beta 1 fuzz targets asserted process robustness and result bounds; they did not prove semantic HTML safety. The v1 compiler target now also asserts deterministic diagnostics and generated Go, valid formatted Go, source-digest binding, bounded public diagnostic shape, and sanitized source map directives. A separate runtime target asserts deterministic, fail-closed URL handling with no partial output. These properties still do not replace the committed differential corpus or real-browser testing.

Assessment findings and remediation status

The 2026-08-12 assessment identified six concrete gaps. Their status in the named public Beta 1 source is recorded here:

Finding	Current remediation	Executable evidence
Generated code named an ABI but did not enforce the exact contract	Generated code now requires the version-specific sando.ABISandoV1 symbol	TestGeneratedCodeRequiresVersionedRuntimeABIMarker; TestRuntimeABIMarker
Deleted or renamed sources could leave owned .sando.go orphans invisible to directory-level check	Directory discovery now reports owned outputs whose adjacent source is absent or non-regular and blocks the operation before writes	TestDirectoryOperationsRejectOrphanedOwnedOutputBeforeWrites
Component-context prose included arbitrary handwritten implementations in the generated-component guarantee	Runtime documentation, policy, and threat model now classify handwritten components as trusted output capabilities	API documentation plus policy review; generated-balance tests retain their narrower scope
An exited development child left its former upstream selected	Exit notification now clears only the matching active target immediately, independent of the watcher poll interval	TestClearTargetOnlyClearsSelectedUpstream; TestSupervisorClearsTargetWhenCurrentApplicationExits
Trusted-value warnings were described more broadly than their analysis supports	Policy and threat-model copy now call them best-effort lexical audit hints rather than type or taint analysis	Documentation assertion and review
Public copy implied a completed systematic html/template differential campaign	Policy and public security copy now describe fixed adversarial cases and list systematic differential work as open	Documentation assertion and review

The exact public Beta 1 source passed the race-enabled repository verifier, sanitized-snapshot tests, both bounded fuzz-smoke targets, compiler/runtime known-vulnerability scans, candidate-version provenance checks, native Windows and executed Linux matrices, and Windows/macOS cross-compilation on 2026-08-12. Signed annotated runtime and compiler tags were then published from that commit in that order. Fresh runtime-first installation passed through both direct Git resolution and the public Go proxy after normal proxy propagation. Future release decisions require fresh evidence for the maintained Linux/amd64 and Darwin/arm64 targets rather than reusing this historical campaign.

Open assurance gaps

• delivery to security@sandwichhime.com is owner-confirmed through a controlled domain catch-all; encrypted reporting, documented backup, and recovery rehearsal remain incomplete;
• the signed annotated Beta tags and their common peeled commit were verified; prebuilt-artifact signing, checksums, SBOM, reproducible provenance, and key recovery remain incomplete;
• Linux/arm64, Darwin/amd64, Windows, and other targets are outside the current maintained release set;
• the exact public candidate still needs the committed real-browser generated document and development-supervisor campaign on both maintained hosts;
• compiler input size, CPU, and memory have no built-in hard budget;
• filesystem checks do not defend against a hostile local actor racing path components between inspection and use;
• the watcher is a convenience mechanism, not a filesystem-integrity monitor;
• human-readable diagnostics can include hostile local filenames or child-tool output and should not be treated as a sanitized log protocol;
• development CSP rewriting is convenience, not production CSP validation;
• deliberately detached child descendants may evade process-tree cleanup;
• rendering has no built-in recursion, output-size, allocation, CPU, panic, or deadline enforcement;
• static cycle detection and trust-use warnings are best-effort analyses; and
• the project has no independent security audit or bug-bounty program.

v1 disposition of open gaps

The list above intentionally mixes incomplete release evidence with boundaries that are not promised by this product. The RC may not convert either category into vague assurance. The following disposition is explicit and remains subject to exact-public-candidate review:

Gap	v1 disposition
Security mailbox delivery, backup, and recovery	Release blocker. Complete the delivery/reply and recovery drill before RC publication. Encrypted reporting may remain optional if the supported confidential channel and its limit are stated accurately.
Artifact signing, provenance, and key recovery	Release blocker. Complete deterministic native artifacts, Developer ID notarization, signed-tag rehearsal, and recovery evidence.
Maintained native matrix	Release blocker for Linux/amd64 and Darwin/arm64 only. Other architectures and operating systems are explicitly unsupported, not silently untested promises.
Real-browser parser and supervisor evidence	Release blocker. The repository-owned gate covers a generated typed document, parsed structure, hostile-value inertness, and supervisor behavior. Execute it against the exact public candidate on both maintained hosts before publication.
Compiler resource budgets	Accepted v1 boundary. The compiler is a trusted local build tool; operating-system and runner limits own CPU, memory, and input quotas. No hostile-input resource guarantee is made.
Hostile local filesystem races	Accepted v1 boundary. Symlinks and ownership are checked, but an actor able to mutate the workspace concurrently is outside the trust model.
Watcher integrity	Accepted v1 boundary. Watching is development convenience; explicit check, generation, Go tests, and builds remain release/deployment authority.
Human-readable child diagnostics	Accepted v1 boundary. They are bounded for resources but remain trusted local terminal output, not a sanitized telemetry format.
Development CSP rewriting	Accepted v1 boundary. It enables reload on trusted loopback pages and is not production CSP validation.
Deliberately detached descendants	Accepted v1 boundary. Ordinary process groups are terminated and waited for; adversarial detachment is outside the trusted-project development model.
Render recursion, output, panic, allocation, CPU, and deadlines	Accepted v1 boundary. Components are ordinary trusted Go; applications own recovery, deadlines, and resource policy.
Static cycles and trust warnings	Accepted v1 boundary. They are documented best-effort audit hints and never replace Go review/tests or explicit trust decisions.
Independent audit and bug bounty	Accepted disclosure, not a security claim. Neither exists for RC. Public tests, threat model, reporting, and correction policy must not be described as an independent audit.

An accepted boundary is permitted only because matching compatibility, threat model, and release copy already avoid the stronger promise. Any conflicting marketing or documentation reopens the item as a release blocker.

Interpreting this ledger

“Pass” means the named command or case produced its expected result in the named environment on the assessment date. It does not mean “secure.” Confidence comes from keeping the boundary small, making risky capabilities explicit, preserving ordinary generated Go for review, publishing reproducible tests, recording failures, and correcting claims when evidence is weaker than the prose.