Harden experiment transport and evidence precision

Author: 0xmrma

.github/workflows/ci.yml

@@ -61,8 +61,23 @@ jobs:
           --cov=mrma.core.sender
           --cov-report=term-missing
           --cov-fail-under=85
-      - name: Runtime dependency audit
-        run: python -m pip_audit --requirement requirements-audit.txt
+
+  runtime-audit:
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.13"]
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Audit published runtime dependency floor
+        run: |
+          python -m pip install --upgrade pip pip-audit
+          python -m pip_audit --requirement requirements-audit.txt
 
   distribution:
     runs-on: ubuntu-latest
@@ -93,7 +108,7 @@ jobs:
           /tmp/mrma-wheel/bin/python -m pip check
           cd /tmp
           /tmp/mrma-wheel/bin/mrma --version
-          /tmp/mrma-wheel/bin/python -c "from importlib.resources import files; root=files('mrma.schemas'); assert all(root.joinpath(f'experiment-v{v}.schema.json').is_file() for v in (2,3,4)); assert 'site-packages' in str(root)"
+          /tmp/mrma-wheel/bin/python -c "from importlib.resources import files; root=files('mrma.schemas'); assert all(root.joinpath(f'experiment-v{v}.schema.json').is_file() for v in (2,3,4,5)); assert 'site-packages' in str(root)"
 
   container:
     runs-on: ubuntu-latest

CHANGELOG.md

@@ -3,6 +3,30 @@
 All notable changes are documented here. MRMA follows semantic versioning for the CLI and uses an
 independent version in each machine-readable evidence schema.
 
+## Unreleased
+
+### Transport Reproducibility
+
+- Experiment transport now disables HTTPX environment trust by default and supports explicit proxy,
+  CA bundle, and environment opt-in policies. V5 evidence records TLS and proxy provenance without
+  exposing credentials, environment values, or file paths.
+- Research and forensic assurance disable environment trust and reject disabled TLS unless the
+  researcher supplies a separate explicit exception.
+
+### Evidence Precision
+
+- Missing, malformed, or conflicting `Content-Type` evidence is no longer assumed to be text.
+  Unequal bodies remain indeterminate unless an explicit weaker assumption is requested.
+- Exact target-specific response fields can be made decision-bearing. Results disclose the full
+  selected header set, omitted-header risk, and a response-header coverage assurance dimension.
+- Experiment evidence advances to strict `mrma.experiment/v5`; published v2, v3, and v4 contracts
+  remain packaged and immutable.
+
+### Verification
+
+- Added transport secrecy, body eligibility, custom-header, schema-negative, and v4 immutability
+  regressions. Runtime dependency auditing now runs independently on Python 3.10 and 3.13.
+
 ## 0.3.5 - 2026-07-14
 
 ### Dependency Contract

README.md

@@ -7,7 +7,7 @@ across a layered HTTP system, then helps reduce that signal to the smallest resp
 It is not a generic vulnerability scanner and it does not treat a one-off response difference as
 a finding.
 
-> Status: `0.3.5` research preview. `mrma experiment` has a conservative evidence contract;
+> Status: `0.3.5` research preview; `main` targets 0.3.6. `mrma experiment` has a conservative evidence contract;
 > legacy survey and minimization commands do not yet share this oracle.
 
 ## The flagship workflow
@@ -52,8 +52,9 @@ The result includes:
 - the effective normalization policy, state and connection modes, response bound, complete retry
   policy, negotiated HTTP versions, and stop reason;
 - a multidimensional assurance profile for statistical decisiveness, control stability,
-  connection independence, state isolation, body completeness, normalization risk, and transport
-  reproducibility, without a scalar confidence label;
+  connection independence, state isolation, body completeness, normalization risk, transport
+  reproducibility, transport integrity, and response-header coverage, without a scalar confidence
+  label;
 - structured limitation codes with severity, scope, explanation, and remediation;
 - a short run ID and evidence schema version.
 
@@ -74,9 +75,9 @@ mrma experiment \
   --fail-on any-signal
 ```
 
-The output declares `mrma.experiment/v4`; exit code `10` means influence and `11` means
-inconclusive when selected by `--fail-on`. The strict v4 JSON Schema defines every nested evidence
-object and cross-field assurance invariant. Published v2 and v3 schemas remain packaged and
+The output declares `mrma.experiment/v5`; exit code `10` means influence and `11` means
+inconclusive when selected by `--fail-on`. The strict v5 JSON Schema defines every nested evidence
+object and cross-field assurance invariant. Published v2, v3, and v4 schemas remain packaged and
 immutable for compatibility. The default exit code remains zero for all verdicts.
 The transport is labeled `semantic-http`; MRMA uses `httpx` and does not claim byte-for-byte HTTP/1
 wire reproduction.
@@ -85,6 +86,9 @@ Responses are streamed with a default 1 MiB read bound. `--body-storage sample`
 KiB per observation. When full normalization cannot be performed, unequal digests are marked
 `INDETERMINATE`; they are never silently treated as equivalent. Encoded and non-text bodies use
 exact transfer-digest equality only until bounded, content-aware decoders are implemented.
+Responses without one unambiguous, well-formed `Content-Type` also use digest-only evidence by
+default. `--assume-text-without-content-type` is an explicit weaker assumption and emits a
+structured limitation.
 
 Retries are disabled by default. When enabled, every intermediate attempt, outcome class,
 error subtype, retry-triggering status, backoff, and final result is recorded. Stable error-subtype
@@ -99,7 +103,10 @@ fingerprint. Response fields use explicit semantics for `Vary`, `Allow`, CORS to
 `Cache-Control`, `Location`, and `Content-Location`. HTTP method tokens remain case-sensitive even
 when their list order is irrelevant. Duplicate directives or malformed `Cache-Control` syntax
 preserve ordered evidence and emit `AMBIGUOUS_CACHE_CONTROL`; captured fields without a registry
-rule and `Set-Cookie` remain conservative ordered evidence.
+rule and `Set-Cookie` remain conservative ordered evidence. The registry is intentionally
+selective: exact target-specific fields can be made decision-bearing with repeatable
+`--include-response-header`; the evidence records the complete selected set and an explicit
+coverage limitation.
 
 Connection scope is explicit: `reuse` (default), `per-arm`, `per-round`, or `fresh-observation`.
 Cookie state remains a separate policy. Wilson intervals assume repeated observations are suitably
@@ -108,8 +115,13 @@ authentication, HTTP/2 state, throttling, or WAF scoring could violate that assu
 
 For publication-oriented work, `--assurance research` authoritatively selects fresh connections,
 isolated response state, disabled retries, a 20-round bracketed design, standard privacy, and full
-body retention within the configured response bound. `exploratory` preserves efficient pooled
-behavior; `forensic` uses research isolation while intentionally retaining exact metadata.
+body retention within the configured response bound. It also disables HTTPX environment trust.
+Experiment transport ignores proxy and CA environment variables by default; opt in with
+`--trust-environment`, or use explicit `--proxy` and `--ca-bundle` inputs. Evidence stores only
+keyed proxy/environment fingerprints and the CA bundle digest, never proxy credentials, environment
+values, or file paths. Research and forensic presets reject `--insecure` unless the separate
+`--allow-insecure-research` exception is supplied and recorded. `exploratory` preserves efficient
+pooled behavior; `forensic` uses research isolation while intentionally retaining exact metadata.
 
 JSON output uses atomic replacement. Add `--evidence-write durable` to flush and synchronize the
 temporary file before replacement and synchronize the parent directory on platforms that support

RELEASE_CHECKLIST.md

@@ -8,7 +8,7 @@
 - [ ] `python -m compileall mrma` passes
 - [ ] wheel and sdist pass `twine check`
 - [ ] clean virtualenv installs the built wheel and passes `pip check`
-- [ ] `pip-audit` reports no known vulnerable runtime dependency
+- [ ] `pip-audit` reports no known vulnerable runtime dependency on Python 3.10 and 3.13
 - [ ] CI passes on Python 3.10 and 3.13 across Linux, Windows, and macOS
 
 ## Versioning
@@ -20,15 +20,19 @@
 - [ ] `mrma config --json` works
 - [ ] `mrma run --url https://example.com --follow-redirects` works
 - [ ] `mrma experiment` detects a deterministic local mutation
-- [ ] `mrma experiment --json` emits schema-valid `mrma.experiment/v4` without decoration
+- [ ] `mrma experiment --json` emits schema-valid `mrma.experiment/v5` without decoration
 - [ ] cookie state does not cross observations in default isolated mode
 - [ ] response limits and transport failures produce typed evidence instead of crashes
 - [ ] redirect and retry traces affect verdicts when final responses are identical
 - [ ] equivalent redirect targets and parsed response headers do not create false signals
 - [ ] method-token case and ambiguous cache-directive order remain decision-bearing
 - [ ] ambiguous cache syntax emits `AMBIGUOUS_CACHE_CONTROL`
 - [ ] retry error subtypes are decision-bearing while timing remains quantitative context
-- [ ] assurance presets, profiles, and structured limitations satisfy v4 cross-field constraints
+- [ ] assurance presets, profiles, and structured limitations satisfy v5 cross-field constraints
+- [ ] experiment transport ignores environment proxy/CA settings unless explicitly enabled
+- [ ] transport evidence fingerprints proxy/CA configuration without exposing values or paths
+- [ ] absent or ambiguous `Content-Type` remains digest-only unless explicitly overridden
+- [ ] exact included response headers affect decisions and selected scope is machine-readable
 - [ ] durable evidence mode syncs the file and the parent directory where supported
 - [ ] every connection mode passes its state and pool-isolation tests
 - [ ] exit codes `10` and `11` match the documented `--fail-on` policy

docs/ROADMAP.md

@@ -54,6 +54,23 @@ adding more mutation families.
 - [x] Signed annotated tag verification in every publishing workflow
 - [x] Protected release environment and verifiable Python/container provenance
 
+## 0.3.5 - Dependency contract consistency
+
+- [x] Align the Python 3.10 Tomli security floor across package and audit metadata
+- [x] Parse and compare published and audited PEP 508 dependency contracts in tests
+- [x] Validate Rich 15 and Actions Checkout 7 across the complete quality matrix
+
+## 0.3.6 - Transport and evidence precision
+
+- [x] Disable ambient HTTPX environment configuration for experiments by default
+- [x] Record privacy-preserving TLS, CA, proxy, and environment provenance
+- [x] Require an explicit exception for disabled TLS under research assurance
+- [x] Treat missing or ambiguous `Content-Type` as digest-only evidence by default
+- [x] Allow exact target-specific response headers into decision evidence
+- [x] Expose selective response-header coverage as a structured limitation and assurance dimension
+- [x] Publish strict `mrma.experiment/v5` while preserving v2, v3, and v4 contracts
+- [x] Audit runtime dependencies on Python 3.10 and 3.13 independently
+
 ## 0.4 - Authorization-first experiment engine
 
 - Enforce an authorization policy before any network operation, including every redirect