feat(mainframe): named call graph and dataset boundary (#3200, #3201)

[squid-protocol]

Closes #3200. Closes #3201.

The counted rules say that a COBOL program calls something and touches files (ipc_rpc_bridges → arch_ipc, io → arch_io). They never say what. This adds the named channel, so the master DB can answer "program P opens DD X for INPUT; job J step S binds DD X to dataset D" — which docs/refraction_engine_differential.md recorded as two stated absences.

Both issues land together because they are one mechanism and one consumer: the skill's ownership table has both gating #3120's lineage switch, so either alone unblocks nothing, and the JCL half of each is the same statement pass (//STEP EXEC PGM= and //DD DD DSN=).

What's new

core/mainframe_boundary.py	extraction: COBOL CALL, CICS LINK/XCTL PROGRAM(...), JCL EXEC PGM=; SELECT ... ASSIGN TO <dd> with the OPEN modes actually used; JCL DD → DSN
core/invocation_resolver.py	repo-wide name → file resolution, and the call/exec edge aggregation
call_site_data	one row per invocation site — resolved or not
dataset_data	the dataset boundary, both the COBOL and the JCL half
edge_data	gains edge_kind 'call' / 'exec'
galaxy_ir.py	dataset_lineage() and unresolved_calls() for the refraction tools

A language opts in with a top-level boundary_extraction declaration (cobol, jcl).

Scores (answer key)

Every engine column that read not carried is now carried, and exact. The forge column is untouched.

corpus	field	before	after
zopeneditor-sample	DD names	not carried	P 12/12 · R 12/12
zopeneditor-sample	inputs	not carried	P 6/6 · R 6/6
zopeneditor-sample	outputs	not carried	P 6/6 · R 6/6
zopeneditor-sample	dynamic CALLs	not carried	P 3/3 · R 3/3
zopeneditor-sample	call targets	not carried	P 3/3 · R 3/3
cics-banking-…-cbsa	DD names	not carried	P 1/1 · R 1/1
cics-banking-…-cbsa	outputs	not carried	P 1/1 · R 1/1
cics-banking-…-cbsa	call targets	not carried	P 45/45 · R 45/45

call targets is a new scored field. It has no forge column: the DAG architect records only non-literal CALL operands and never sees EXEC CICS LINK/XCTL — 140 of CBSA's 144 call sites.

The extractor was scored directly against the key before being wired in: 147/147 call sites match on verb, form, operand, target and line, and 13/13 dataset records on internal name, DD and modes, with no false positives. JCL has no answer key, so EXEC PGM= was checked against an independent raw-file scan: 24 / 77 / 150 steps on the three corpora, exact.

Verification

• Golden master: PASS in both modes, zero diff. No bless owed — no counted signal changed and none of this reaches the audit JSON.
• Refraction snapshot: unchanged, all three corpora, excerpts and full (45/192/259 files, 0 differ), which is what an engine-side PR requires.
• ruff / mypy / dead-key audits clean; ruff format back at the pre-existing 29.
• ReDoS: every new pattern scales linearly (~4× for 4× input, 40k → 160k), no offender under the repo's own rule.
• Delta parity: a full then an incremental scan of zopeneditor produce byte-identical call_site_data and dataset_data.

The one mypy baseline line moved 386 → 388 (pure line shift from two new imports); the message is unchanged.

Three decisions worth reviewing

1. A call edge is not a dependency edge. edge_kind is 'call'/'exec' and these never enter the DiGraph, so pagerank_score, popularity, internal_dependency_links, betweenness, archetypes and every risk score are byte-for-byte unchanged. Whether a runtime invocation should count as coupling is a scoring question with its own before/after — filed as Decide whether call/exec edges should feed the dependency graph (pagerank, popularity, blast radius) #3237, deliberately not settled here. Consequence: Persist the dependency graph edge list (enables neighborhood/interaction-level risk hypotheses) #2992's per-file reconciliation is now scoped to WHERE edge_kind = 'import', and so is galaxy_ir's copy_deps.
2. A CALL resolves by PROGRAM-ID, nearest-wins. That is not the import resolver's rule, which refuses to guess on an ambiguous stem (Dependency resolver drops ambiguous targets: COBOL COPY loses most copybook edges (same stem as the program, or duplicated copybook roots) #3199). An import names a file; a called program is chosen by library concatenation order at link-edit time, which is what the answer key records for zopeneditor's two SAM2 programs. target keeps the name regardless, so nothing is lost when the file attribution is debatable.
3. Unresolved is data, not a gap. Most real sites resolve to nothing — CALL 'CEEGMT' is an LE service, EXEC PGM=IEFBR14 a system utility. The table distinguishes "the name itself was unreadable" (target IS NULL) from "named but external" (dst_file_id IS NULL).

Out of scope, stated

• FD/01 record layouts (Extract COBOL SELECT/ASSIGN and OPEN modes (and JCL DD bindings) by name, so dataset lineage can come from the engine #3201's third bullet): that is data-division item extraction — the differential doc's separate "no data items are extracted" absence — and needs a level-number/PIC/OCCURS/REDEFINES walker. Lineage does not need it.
• Reachability: an OPEN in an unreachable paragraph is still extracted. The engine has no reachability model, and inventing one here would repeat the mistake cobol usage_status: entry paragraph flagged, case-sensitive matching, and NAME-EXIT counts as a reference to NAME #3198 corrected.
• The forge-side switch is a separate PR, per the skill's "don't bundle them".

Two traps, recorded because each nearly shipped

• The orphaned_logic reports 100% dead code for five languages whose invocation model never names the callee — the family #2727 closed without covering #2806 string-in-rules trap. The declaration was first written as a string inside rules, which language_lens.py silently re.compile()s. Every unit test stays green while a real scan extracts nothing. It is now top level, with a test asserting it never moves back.
• A leaked sys.modules entry (fixed here). Two tests in test_galaxyscope.py injected a MagicMock over gitgalaxy.core.state_rehydrator and never restored it, so every later test in the session that imported StateRehydrator got a mock whose load_state returns {"commit_hash": "old", "ram_cache": {}}. The new delta tests pass alone and in their own directory, then failed in the full suite with an empty cache and a KeyError pointing nowhere near the cause. Restored with addCleanup at both sites — pre-existing pollution, but this PR is the first thing to trip it.
• A verb inside a string literal. DISPLAY 'GNP CALL FAILED :' read as a call to FAILED — 5 of 102 sites on carddemo, which has no answer key and so was audited separately. Neither scored corpus contains the shape. Guarded per line (quote state never crosses a line, or AUTHOR. James O'Grady. swallows the file — the COBOL modernization answer key: hand-verified per-program truth for the #3120 corpora, scored against forge and engine #3210 trap), and pinned.

Filed along the way

• function_data.calls_out_to is non-deterministic: list(set)[:20] keeps a different subset per PYTHONHASHSEED #3236 — function_data.calls_out_to is list(set)[:20], so it keeps a different subset per PYTHONHASHSEED. Pre-existing and untouched here (detector.py has no diff on this branch), but it was the only non-additive difference between a pre- and post-change scan and had to be chased down to prove this PR moved nothing.
• Decide whether call/exec edges should feed the dependency graph (pagerank, popularity, blast radius) #3237 — the deferred graph question above.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab