docs(assessment): define external proof boundaries

docs/CAPABILITY_AND_INFRASTRUCTURE_FIT.md

@@ -372,6 +372,103 @@ metadata, published-keyring integrity and optional local tag provenance; it does
 not prove installed artifacts, target providers, target infrastructure, or
 production fitness. Those remain separate proof checks above.
 
+### Installed-composition evidence
+
+The same rerun can inspect the Python environment in which it executes and bind
+that observation to the assessment and signed catalog:
+
+```bash
+./.venv/bin/python tools/assessments/capability-fit.py \
+  --public \
+  --trusted-keyring /srv/govoplan/trust/catalog-keyring.json \
+  --collect-installed-evidence /var/tmp/govoplan-installed-evidence.json \
+  --output /var/tmp/govoplan-fit-review.json
+```
+
+The collector follows the strict
+[`installed-composition-evidence.schema.json`](installed-composition-evidence.schema.json)
+contract. It enumerates all installed distributions whose normalized name starts
+with `govoplan-`, compares the enabled assessed package and module-manifest
+versions, and identifies missing, duplicate and extra GovOPlaN distributions.
+Those differences produce stable review targets for the affected composition
+entries and evidence-backed conclusions.
+
+For each distribution, the collector also:
+
+- verifies every supported SHA-256 entry in installed wheel `RECORD` metadata,
+  within fixed limits of 256 GovOPlaN distributions, 10,000 files and 512 MiB
+  hashed per distribution, 50,000 files and 2 GiB hashed for one collection,
+  and 64 MiB for any single file;
+- distinguishes immutable VCS commits and archive hashes from editable installs,
+  local directories, package-index/unknown origins and malformed metadata;
+- compares a non-editable VCS commit with the assessment commit and, when one is
+  present, the signed catalog `selected_units` commit;
+- records only package/module identifiers, versions, bounded counters, hashes
+  and stable error codes. It never writes direct URLs, paths, hostnames,
+  usernames, exception text or file contents.
+
+`RECORD` agreement proves that the files agree with their installed metadata; it
+does not make self-consistent metadata a trusted release origin. Similarly, a
+version string is not artifact integrity. Editable and directory-backed installs
+remain mutable even when their small editable-install `RECORD` is valid. Archive
+or index artifacts without a hash anchored by the assessed release remain
+unanchored. These conditions cause review rather than being promoted to
+immutable installed-release proof.
+
+Use `--installed-evidence PATH` to review evidence collected in a separate
+environment. Evidence input and output are bounded to 4 MiB. Collection loads
+the `govoplan.modules` entry-point factories in order to read module IDs and
+manifest versions; that executes installed GovOPlaN manifest code. Run it only
+inside the installation being assessed and with the same isolation expected for
+other installed-artifact acceptance checks. This collector does not observe
+which modules a running service activated, migrations, configuration, health,
+or reference-journey behavior. Runtime activation therefore remains an explicit
+unchecked boundary.
+
+### Target, provider and production proof boundary
+
+Installed evidence cannot approve a target environment, an external provider,
+or production use. These scopes use a separate, expiring
+[`capability-fit-boundary-evidence.schema.json`](capability-fit-boundary-evidence.schema.json)
+bundle. The bundle is bound to the assessment ID, assessment release and exact
+installed-evidence SHA-256 digest. It contains only opaque subject/control/result
+IDs and content hashes, not endpoints, credentials, people or raw result files.
+
+Boundary evidence is accepted only when at least one Ed25519 signature validates
+against a separately provisioned
+[`capability-fit-proof-authority-keyring.schema.json`](capability-fit-proof-authority-keyring.schema.json).
+Each authority key explicitly lists the scopes it may attest. Target and provider
+claims use `passed` or `failed`; production claims use `approved` or `rejected`.
+One claim per scope, unique control/artifact IDs, `issued_at < expires_at`, current
+validity and exact digest binding are mandatory. Any schema, binding, time,
+signature or authority blocker leaves every supplied boundary claim unchecked;
+an authorized negative result is checked but requires review.
+Signatures cover UTF-8 JSON with the `signatures` member omitted, object keys
+sorted, compact `,`/`:` separators and non-ASCII characters escaped, matching
+the tool's deterministic canonicalization.
+
+```bash
+./.venv/bin/python tools/assessments/capability-fit.py \
+  --public \
+  --trusted-keyring /srv/govoplan/trust/catalog-keyring.json \
+  --installed-evidence /srv/govoplan/evidence/installed.json \
+  --boundary-evidence /srv/govoplan/evidence/target-proof.json \
+  --boundary-authority-keyring /srv/govoplan/trust/proof-authorities.json
+```
+
+The authority keyring is a governance trust root. Do not download or generate it
+from the proof bundle being checked, and do not reuse catalog signing authority
+as implicit production authority. A target operator or approver must validate
+the referenced drill/result artifacts before signing. The rerun verifies the
+attestation and its bindings; it does not fetch or reinterpret those artifacts.
+`--verification-time` exists for reproducible historical review. Live admission
+must use the actual current time.
+
+No boundary bundle or production authority has been supplied for this current
+assessment. Target environment, provider and production proof therefore remain
+explicitly unchecked rather than inferred from the local GreenMail journey,
+source tests or signed release metadata.
+
 ## Evidence used in this slice
 
 - [Production-like profile](../dev/production-like/README.md) and
@@ -389,6 +486,9 @@ production fitness. Those remain separate proof checks above.
   verified against a separately provisioned local trust keyring
 - Annotated source tags `govoplan-core/v0.1.13` and
   `govoplan-campaign/v0.1.10`, including their catalogued Python and WebUI refs
+- Installed-composition, boundary-proof and independently scoped proof-authority
+  schemas plus their deterministic review tests; no current target or production
+  proof bundle is asserted
 
 Checks retained from the first assessment slice against its then-current
 workspace:

docs/capability-fit-current.json

@@ -967,10 +967,12 @@
   ],
   "proof_checks": [
     "Materialize the signed catalog into an isolated installation and rerun contract, migration and module-permutation gates against the installed artifacts.",
+    "Collect the isolated installation with the bounded installed-composition evidence contract; require exact enabled package/module versions, complete RECORD verification and immutable provenance anchored to this assessment.",
     "Run a safe target-like Campaign through SMTP acceptance, IMAP append, reporting and audit.",
     "Drill worker, Redis and ambiguous-delivery failures without duplicate sends.",
     "Restore PostgreSQL, managed files, configuration and encrypted credentials and measure RPO/RTO.",
     "Validate proxy/TLS, cookies/CORS, account bootstrap, secret redaction, monitoring and alert delivery.",
-    "Measure representative Campaign/file/queue/database load and external throttling."
+    "Measure representative Campaign/file/queue/database load and external throttling.",
+    "Require separately issued, expiring and independently scope-authorized evidence before marking target environment, external provider or production approval proof as checked."
   ]
 }