mimosa 1.0.3: guard shapes not credited by taint analysis (5 reproducible defects) + request incremental git-gate

[devighd-arch]

Mimosa upstream feedback draft (M1)

Status: drafted, pending submission. Once submitted, put the tracking id into
mimosa_triage.json → upstream.tracking.

Environment

• Plugin: mimosa@zcode-plugins-official 1.0.3
• Workload: monorepo with a live trading system (~370 source files, py/ts)
• Baseline seals: scan-2026-09-28T04-25-50.259Z-22d8d1a1af55 (62 findings),
  scan-2026-09-28T04-39-14.508Z-223908513145 (32 findings post-fix)

Five reproducible defects / gaps

1. Sink-wrapped guard not credited. chainlink_poll.py already called
   guard_url(url) at URL construction (landed in a reviewed commit); its
   urlopen kept being flagged. Moving guard_url(...) inline into the
   Request(...) constructor at the sink statement still does not clear the
   finding. Expected: a call to a guard function wrapping the URL argument at
   or near the sink should satisfy the SSRF rule (scheme+host+resolved-IP are
   all enforced by the guard).

2. Same-shaped fix clears one file but not another. A function-entry
   urlsplit scheme+hostname allowlist cleared pmracer/pmracer/history.py
   and pmracer/scripts/probe_sources.py but not
   pmracer/scripts/backfill_candles_35d.py (byte-for-byte same shape).
   Suggest a documented, deterministic recognition contract for guard shapes.

3. Helper-mediated guards are opaque. _safe_join() (containment via
   os.path.realpath + commonpath) results still get flagged at every
   downstream open(). Same for cross-module guard_url. Expected: local or
   cross-file sanitizer functions be recognized (even via an explicit
   allowlist of guard function names configured per project).

4. Git gate scans the wrong tree. Committing in C:\repo-dev while the
   ZCode workspace is C:\repo made the L3 gate report findings with
   C:\repo\... paths (stale live tree), blocking every dev-tree commit until
   a subprocess bypass was used. Expected: resolve the tree from the git
   command itself (-C, cd chains) or the tool payload cwd, and scan only
   the committing tree; ideally scan only the commit's staged diff
   (focusFiles) instead of the whole project.

5. Scan runs are frequently inconclusive via node spawnSync ETIMEDOUT.
   Five consecutive stop-hook runs in .mimosa/history are inconclusive
   with spawnSync <node.exe> ETIMEDOUT on core files (tail_engine,
   shadow_engine, accounts). Expected: configurable timeout, retry, and an
   explicit verdictEffect policy when the scanner times out.

Requested sanitizer recognition (would clear our remaining 23 high)

Pattern id	Shape	Files
sink-inline-guard	urlopen(Request(guard_url(url), ...))	planb feeds/ledger/truth/tail_engine/context (7 sites)
entry-allowlist	urlsplit scheme+host allowlist at helper entry	backfill, datastreams probe, price_history
inline-dotdot	explicit '..' in path.parts (or string split) before open()	identity, calibration, counter_report, segment_analysis_v2
safe-join-inline-dotdot	_safe_join() + re-stated inline check at sink	regime_lab, segment_analysis

The seeded-randomness rule also fires on random.Random(seed) used for
statistical simulation (bootstrap / CRN Monte-Carlo); those sites cannot move
to secrets without breaking reproducibility. A project-level rule tweak or
a # mimosa: seeded-simulation annotation contract would help.