map: match full input paths and bind flows to the exact sink call

[patchstackdave]

The two blockers an external review named before map can be used as an automatic rule-authoring source. Both let precise — the signal a generator would pin a parameter on — identify the wrong input.

1. Path collisions

Matching reduced both inputs and reads to their last segment, so billing.email and shipping.email both became email: reading one marked both precise.

Flows now compare canonical full paths (data.shipping.email → shipping.email), with index tokens erased so tags[0].label, tags[].label and tags.label compare equal. precise requires the whole path to match, or the input to be an ancestor of what was read (billing legitimately covers billing.email).

2. Sink ownership

Evidence was gathered from the enclosing statement, so a sibling expression could lend it:

await Promise.all([ audit(data.title), db.from("items").insert({ title: "system" }) ]);

…reported title → insert [precise] although the insert only receives a literal.

Sinks now carry their source span (start/end) as their identity, and evidence comes only from that call's own arguments plus other calls in the same fluent chain (.update({…}).eq('id', data.id) is one operation). Walking up stops at an array literal or argument position — which is exactly what excludes the sibling.

Two subtleties worth noting: a line is not an identity, and neither is a start offset alone — in db.from(t).insert(x) both calls start at db, so the key is start:end.

Two recall bugs found while testing

Both silently produced zero precise flows for the most common shapes:

• A validated payload wasn't tainted. Schema.parse(await req.json()) broke the chain, so every validated handler lost its flows. Validation is not sanitization — a validated value is still attacker-controlled — so taint now survives parse/safeParse.
• A destructured handler param was treated as a path segment. { data } (TanStack's validated-payload convention) shifted every path by one segment, killing all matching.

Result on the real reference app

More accurate than before — the spurious helper-reached select flows are gone, and the chain-fed filters are correctly precise:

createTask:   precise=[title->insert]
toggleTaskFn: precise=[id->update, completed->update]
deleteTask:   precise=[id->delete]
getTasks:     precise=[]

Sinks also expose start/end in the emitted map, as the review asked, so a consumer can point at the exact call. 784 tests green; typecheck clean.

Not in this PR

The review's larger rule-generation roadmap — argument roles (command vs args, URL vs body, path vs contents, raw SQL vs values), the finer confidence taxonomy (exact/transformed/interprocedural/heuristic/unknown), a TypeChecker-based dataflow engine, explicit why-not-generatable reporting, the candidate lifecycle, and corpus metrics. Those are the next tranche, tracked separately; this PR only removes the two correctness blockers.

[coderbuds]

Precise span-based flow binding greatly improves path matching accuracy.

🎯 Quality: 80% Excellent · 📦 Size: Large — consider splitting if possible

🛡️ Standards: no pre-flight fit check ran for this change — wire assess-change-fit into your coding agents to catch size before opening.

📈 This month: Your 56th PR — above team average · Averaging Good

_{See how your team is trending →}

[patchstackdave]

/review