C4b: allocation accounting, fetch/exec byte caps, and --deadline

[howlcipher]

Summary

Implements review candidate C4b on top of C4a (#111 / cdf2e2fc): bytecode runner resource limits for S4/S6/S7.

• Allocation accounting against MaxMemoryBytes (default 64 MiB, --max-memory-bytes). Cumulative approximate charges on string concat, append, MAKE_LIST/MAKE_DICT, str_split/str_join, read_file, fetch/exec bodies, encode_json, and related large sites. Shared counter across SPAWN_AGENT children. Exceed → structured LIMIT_EXCEEDED.
• Fetch body / exec output caps (default 10 MiB each; --max-fetch-bytes, --max-exec-output-bytes). LimitReader / bounded combined stdout+stderr; fail-closed for non-positive ceilings in the VM; CLI rejects <= 0.
• --deadline optional positive time.ParseDuration wall-clock budget on ExecutionPolicy. Threaded context.Context into fetch, exec, sleep, and model HTTP calls; instruction loop also checks cancellation. Absent flag = no deadline; explicit 0s/invalid rejected.
• CLI wired on both legacy -run-bc and howlframe run (bytecode/bc), alongside existing --max-instructions / --max-call-depth.
• Docs: change_log.md, journal docs/journals/2026-10-06_c4b_resource_limits.md, doc 11 C4 status, S4/S6/S7 honestly marked PARTIAL where AST/print/read_line remain open.

Hard nos intact

No prod -compile-bc/HFIR flip; #90 stays Partial; no HFBC/opcode change; no DOM invent; no C5 receipts/harness; C4a MaxCallDepth behavior preserved.

Defaults

Flag	Default
--max-memory-bytes	67108864 (64 MiB)
--max-fetch-bytes	10485760 (10 MiB)
--max-exec-output-bytes	10485760 (10 MiB)
--deadline	absent (none)

Deferrals

AST interpreter memory/deadline/recursion still open; print uncapped; blocking read_line (and some host I/O) not deadline-cancelled; charges are approximate construction accounting, not exact RSS.

Test plan

• gofmt -l . clean
• go vet ./...
• go test ./internal/vm/ -count=1 (incl. new resource limit / deadline / byte-cap tests)
• go test ./... -count=1 green outside sandbox (Go 1.24.4)
• C4a call-depth / instruction-limit CLI regressions still pass

[howlcipher]

Dev-lead COMMENT (head 9f55bc65aeee607b09276422f6faab9a4355e297)

Verdict: merge-ready. Diff matches claimed C4b scope. CI build green. Undraft + squash-merge when you are ready.

What matches

• Memory: shared allocationBudget (mutex, never reclaimed) charged at concat/convert strings, MAKE_LIST/MAKE_DICT, APPEND, STR_SPLIT/JOIN, map_set new keys, map_keys slots, encode_json, HTML escape, read_file/fetch/exec via chargeBytes (source + 8× slot ≈ 9n), and model response strings. Overflow-safe chargeSlots. Fail-closed for MaxMemoryBytes <= 0 on positive charges. Shared into SPAWN_AGENT, legacy SPAWN, and HTTP handler children.
• Fetch/exec caps: default 10 MiB each; LimitReader(max+1) then reject; limitedOutput Write-only (no ReadFrom bypass) + cancel on overflow; under-cap nonzero exit stays IO_ERROR. CLI + VM fail-closed for <= 0.
• Deadline: optional ExecutionPolicy.Deadline; WithTimeout shared with children; instruction-loop checkDeadline; sleep select; fetch/exec/model via context; deadline → structured LIMIT_EXCEEDED (not IO_ERROR). Absent = none; CLI rejects 0s/invalid.
• CLI: --max-memory-bytes / --max-fetch-bytes / --max-exec-output-bytes / --deadline on -run-bc and run (bytecode), via resourcePolicy. C4a --max-call-depth / MaxCallDepth preserved.
• Tests: defaults, opcode boundaries, spawn share, fetch/exec caps (httptest + in-memory transport), deadline sleep/loop/exec/fetch/model/response-body, CLI both runners, inherited exec without deadline unchanged.
• Docs: S4/S6/S7 honestly PARTIAL; deferrals (AST, uncapped print, read_line not cancelled, approximate ≠ RSS) match the code.

Non-blockers (documented)

• Append charges full list size each time (cumulative overcharge vs live heap) — intentional approximate policy.
• Model response bodies have no dedicated byte cap (only alloc charge + request deadline).
• Deferred paths as claimed: AST interpreter, print, blocking read_line / some host I/O.

Hard nos

Intact: no prod HFIR / -compile-bc flip, #90 Partial, no HFBC/opcode change, no C5/harness/DOM; C4a call-depth behavior preserved.

No blockers.