feat(http): record/replay for net.fetch (closes #7)

[escapeboy]

Sprint 0.5-S7 — closes #7 (FleetQ P2, fifth in a row)

Boruna scripts are deterministic by design; external HTTP is not. This bridges the gap so agent CI loops become genuinely reproducible. Distinctive selling point per FleetQ: "one of the few runtimes where record/replay would be ergonomic." Pulled forward from 0.5.0 (same pattern as #6, #8).

Surface

```bash

Record once against the real upstream:

boruna run app.ax --policy allow-all --live \
--record-net-to fixtures/run-001.tape.json

Replay forever, no network access:

boruna run app.ax --policy allow-all \
--replay-net-from fixtures/run-001.tape.json
```

Tape file:
```json
{
"format_version": 1,
"transactions": [
{ "method": "GET", "url": "https://api.example.com/users/42",
"request_body": null, "response_body": "{\"id\":42}" },
{ "method": "POST", "url": "https://api.example.com/events",
"request_body": "{\"event\":\"click\"}",
"response_body": "{\"ok\":true}" }
]
}
```

Match strategy (locked, in design doc)

• Strict ordered, key on `(method, url, request_body)`
• Headers EXCLUDED from match key — auth tokens change between sessions
• Mismatch → typed error: `position N: method differs / url differs / request_body differs`
• Exhaustion → typed error: `tape exhausted (N consumed, asked for more at position N)`
• Under-consumption → silently OK (trailing tape entries unused)

Tests

• 18 new tests in `boruna-vm` (round-trip, in-order match, all three mismatch flavors, exhaustion, under-consumption, default-method-GET, case normalization, mock pass-through, save-on-drop, drop-without-path, parser-agreement, back-to-back-identical-calls for polling scripts, format-version compatibility, save-then-load round trip)
• All 124+ existing VM tests pass under `--features http`
• All 591+ existing workspace tests pass
• `cargo clippy --workspace -- -D warnings` clean (with and without http)
• `cargo fmt --all -- --check` clean

Review

`ce-correctness-reviewer` surfaced 1 HIGH + 3 MEDIUM findings. All addressed before commit:

#	Finding	Fix
1 (HIGH)	Save-on-drop swallows tape errors → CI fixture corruption	CLI write-probe at startup; Drop is the safety net only
2 (MED)	Tape lost on panic mid-recording	Documented as v1 limitation; streaming tape is the future fix
3 (MED)	Missing test for back-to-back identical calls (polling)	Added
4 (MED)	`describe_net_fetch_request` could drift from `HttpHandler` parser	Extracted shared `parse_net_fetch_args` in `http_handler.rs`; both call it. Regression test asserts agreement.

Documented limitations (per review)

• Request headers NOT in match key (auth tokens change between sessions)
• Response status/headers NOT recorded (handler returns body only today)
• Failed transactions NOT taped in v1
• Non-UTF-8 response bodies inherit `HttpHandler`'s "not valid UTF-8" error
• Tape file size unbounded (~2× pretty-print multiplier)
• Panic during record may lose the tape (esp. `panic = "abort"`)

What's NOT in this PR (follow-ups)

• `boruna workflow run --record-net-to / --replay-net-from` — same machinery; defer to follow-up sprint when there's a workflow-level ask
• `boruna_run` MCP parameter form — defer until asked
• Recording response status + headers — requires `HttpHandler::handle_net_fetch` return-type change; defer
• Recording failed transactions — requires teaching the tape format about errors; defer
• Header-aware matching — opt-in `match_headers: [...]` config; defer
• `db.query` / `llm.call` record/replay — each is its own sprint; this PR establishes the pattern

Closes

• Closes [P2] Record/replay for net.fetch #7 (FleetQ P2: record/replay for net.fetch)

FleetQ status after this PR

5 of 9 P1/P2 asks closed (#3, #5, #6, #7, #8). Only #9 (per-call OpenTelemetry observability) remains as the last small-sprint pick before the big-sprint pivot to `0.3-S2` (persistent state).

🤖 Generated with Claude Code