Tensor Grep Index Fingerprint Freshness

tensor-grep: index fingerprint + reuse freshness

An index is a memory of a PAST tree. Serving it on a LATER query is only correct if the past tree is still the queried tree — and "still the queried tree" is a WEAKER claim than "nothing changed", because the same path string can name a DIFFERENT tree (a swapped symlink/junction, a renamed tree, a copied/tampered .tg_index). The M17 finding (index reuse) was the exact case: staleness_reason checks only no-ignore mode + per-entry mtime/size + a new-file walk over self.root — it never asks "is the stored root the queried root, canonically?".

Part 1 — Identity-anchor-first: canonical root, never a raw path string

The M17 structural fix (merged on origin/main as v1.110.12; current tag v1.110.16 (re-derive with: git describe --tags origin/main)) is the model:

• Persist a canonical root at build: canonical_root_of() recorded once (format v6) — lexical AND canonical, so later path-spelling differences (case, .., junctions in the prefix) cannot manufacture a mismatch that is not real.
• Store entries canonical-relative to that root with a single deref (canonical_root.join(rel)), so NOTHING is cwd-dependent — a later query from a different working directory must not change where entries resolve.
• At reuse/warm-load, compare canonical QUERY root vs canonical STORED root at EVERY load site — the preloaded_index arm handed off by detect_warm_index_state AND the fresh-reuse branch — and the check must run BEFORE staleness_reason/incremental update.
• On mismatch: REBUILD from the current tree (always safe in-tree — prefer rebuild over bare refuse) with a disclosed reason — rebuild must not be invisible latency.

Root bytes are already persisted (index.rs, canonical_root field — grep the symbol). Two corrections to an earlier pass (2026-08-12): the field is canonical_root, NOT root (the pre-fix plan this skill was drafted from used a bare root field; format v6 persists the canonical form, declared canonical_root: PathBuf); and the plan's line refs copied into that pass (:272-274 write / :355-356,413 read) were the PLAN's own pre-fix numbers, unlabeled — on this tree the write is the format-v6 serializer emitting canonical_root_bytes and the read is the deserializer parsing it back (re-grep canonical_root in rust_core/src/index.rs to relocate both). The defect was that the reuse path never COMPARED them to the query root.

Part 2 — The alias/swap RED shape (the repro that actually bites)

CRITICAL constraint: resolve_index_path puts .tg_index INSIDE the tree, so a plain "build root A, query root B" NEVER shares an index file — the naive RED is unreachable. The real wrong-tree serve needs an ALIAS: same string, swapped tree:

• RED fixture = a root symlink/junction under a FIXED path, swapped to a different directory BETWEEN build and query (Windows: a junction — see tensor-grep-cross-platform-path-confinement Part 1); on a metadata-preserving swap this is the ONLY shape that still looks fresh.
• The fixture carries the fixture-BITES precheck (prove the swap actually applied — the junction must resolve to the NEW dir before the probe runs; see tensor-grep-hermetic-hostile-tests).
• A naive stored_root == query_root string compare PASSES on the alias (same string) — the canonical-root comparison is what separates "same text" from "same tree".
• Keep the per-file identity walk (staleness_reason) as the "still the queried tree" backstop — it is not replaced, it is supplemented.

Part 3 — The fingerprint ladder: climb only as far as the threat demands

Receipts that shaped the ladder: the M17 gate round-2 finding that the initial 4 KiB-per-file byte cap let a same-size/same-mtime edit past byte 4096 in a SAMPLED file evade detection, and that inode identity would not catch it either — so sampled files are hashed in FULL (bounded by the 32-file cap, not by bytes). See compute_tree_fingerprint (index.rs, grep the symbol).

Part 4 — Bounded sampling with NAMED boundaries (the reopen-trigger discipline, A49)

compute_tree_fingerprint samples the top-32 top-level FILES (direct children) of the canonical root — the walk filters to regular files BEFORE the cap (entry.file_type()...is_file() precedes .take(TREE_FINGERPRINT_TOP_LEVEL_CAP)), so a directory or symlink can neither flip the digest nor consume a sample slot — fully hashed, index-machinery namespace excluded (a leftover ..tg_index.<token>.tmp atomic-save temp / ..tg_index.lock must not consume a sample slot — is_tg_index_owned_entry, whose matcher is starts_with("..tg_index.") — TWO leading dots, not one). The honest remaining boundary is NAMED, not hidden:

• The cap (TREE_FINGERPRINT_TOP_LEVEL_CAP = 32) is documented AT the constant with the reason ("deliberately bounded in FILE COUNT; each sampled file is FULLY content-hashed").
• The files NOT sampled (33rd+ top-level files, and every NON-top-level file) are covered only by the per-file mtime/size loop — that gap is M17-FU1, with owner + reopen trigger, recorded canonically (see docs/plans/2026-08-08-backlog-completion-plan.md M17 section and BACKLOG's M-follow-up rows).
• A sampling cap is an honest tradeoff, but it must be a named tradeoff with a reopening trigger (A49), never a silent assumption — "a same-path metadata-preserving swap landing in an unsampled file that mtime/size cannot detect" is that trigger.

The M17 REOPEN TRIGGER pattern applies to every fingerprint surface in the repo, including the Python parse/semantic caches:

• semantic_index.py stale check = SHA-256 fingerprint over sorted paths + mtimes; on mismatch it warns to stderr and falls back in-memory (NOT wired to a tg index command yet).
• repo_map.py _mtime_aware_cache (grep the symbol) is the Python cache-freshness wrapper with a _MTIME_CACHE_CLEAR_REGISTRY that sweeps every decorated cache — and the per-root keyed re-export/tsconfig contexts ARE swept too, contrary to an earlier pass of this skill that said they were "deliberately NOT" among them: _clear_all_source_caches (grep the symbol in repo_map.py) explicitly clears _JS_TS_REPO_CONTEXTS, _RUST_REPO_CONTEXTS, and calls lang_go.clear_go_repo_context_cache() with the comment "Clear them too so the sweep is actually complete" — they are not _mtime_aware_cache wrappers, so the sweep adds them by hand.

External anchors (Exa research, 2026-08-09)

Repo receipts to cite by symbol, not line: index.rs root / staleness_reason / compute_tree_fingerprint / TREE_FINGERPRINT_TOP_LEVEL_CAP / is_tg_index_owned_entry; main.rs resolve_index_path / the reuse branch / detect_warm_index_state; src/tensor_grep/core/semantic_index.py (the SHA-256 mtime-path fingerprint + fallback); repo_map.py _mtime_aware_cache / _MTIME_CACHE_CLEAR_REGISTRY; AGENTS.md A44/A49 and the verification-oracle family (a mismatch that is not real is a Form-10 branch-unit false green). Grep the symbol, never a hardcoded line number; and record every deferred fingerprint boundary as a named M-FU row with owner + reopen trigger.

Quick reference

[1] identity   canonical query root vs canonical STORED root, at BOTH load sites, BEFORE staleness
[2] REPRO       wrong-tree serve needs an ALIAS/SWAP (junction), never the naive wrong-path RED
[3] ladder      mtime+size -> ctime -> full-content checksum -> whole-tree; climb to the threat
[4] named cap   32-file sample is bounded AND the unsampled gap is M17-FU1 with an owner+trigger
[5] mismatch    rebuild with a disclosed reason, never invisible latency, never bare refuse

The endpoint: an index that never serves a tree it was not built for, whose freshness proof is as strong as the threat requires, and whose sampling cap is an honest, named, reopenable boundary.