file-safety (S2/S4 follow-up): track read completeness in FileObservation; require complete observation for full-file update

[easonLiangWorldedtech]

Scope

Follow-up from the CodeRabbit review of #1408 (S4b, epic #1375 file-safety series). The review finding (comment on src/core/tools/ReadFileTool.ts:238) points out that read completeness is not tracked:

• processTextFile() is slice-based by default: even a read_file without explicit offset/limit returns at most DEFAULT_LINE_LIMIT lines (truncated, with a warning), and indentation mode returns only a semantic block.
• The observation registered after such a read is file-level only — FileObservation = { version, observedAt } (src/core/task/observationRegistry.ts), with the version token derived from on-disk fs.stat.
• So fs.readFile can complete successfully, the version tokens can match, and a later write_to_file full-file replacement still passes the guardedWrite() update path — even though the model only ever saw a slice of the file.

Proposed behavior

1. Extend FileObservation with read scope (S2 registry, feat(task): per-task file observation registry (A2, #1375) Zoo-Code-Org/Zoo-Code#1394): e.g. complete: boolean plus the observed line range / total line count (from processTextFile's result: returnedLines, totalLines, includedRanges).
2. Registration records the scope (ReadFileTool): a read that delivered the whole file (or an explicit full-file read) marks the observation complete; slice/indentation/truncated reads mark it partial with the delivered range.
3. guardedWrite() requires a complete observation for the update kind (full-file replacement): a partial observation rejects the update with the standard stale/guard rejection (re-read remediation), while the edit kind keeps working on any prior observation (optionally, a later step could also scope edit to the observed range).
4. Re-reading the file extends/refreshes the observation as today; a complete read supersedes a partial one.

Acceptance criteria

• Regression test: file read through a truncated slice → write_to_file full replacement → rejected with the standard guard error; after a complete read → allowed.
• Regression test: edit kind still permitted after a slice read (no behavior change for scoped edits in this step).
• FileObservation scope fields round-trip through the registry semantics (re-observe replaces the entry).

References

• feat(tools): wire guarded writes into the diff-view save paths (S4b, #1375) Zoo-Code-Org/Zoo-Code#1408 (S4b diff-view wiring) — review finding thread on ReadFileTool.ts:238.
• [EPIC] File Write Safety Prevent Concurrent Write Races Data Corruption Zoo-Code-Org/Zoo-Code#1375 (epic) — A1 version tokens / A2 observation registry / A3 CAS guard.
• feat(task): per-task file observation registry (A2, #1375) Zoo-Code-Org/Zoo-Code#1394 (S2) — observation registry; feat(tools): guarded write CAS core with per-path FIFO chain (S4a, #1375) Zoo-Code-Org/Zoo-Code#1405 (S4a) — guardedWrite() core; feat(tools): wire guarded writes into the diff-view save paths (S4b, #1375) Zoo-Code-Org/Zoo-Code#1408 (S4b) — tool wiring.
• Fork sibling issue #44 — the other 1408 follow-up (diff-view user-accept save path).