fix(detectors): stop the context detector firing on ordinary prose

[mazzasaverio]

Stacked on #57, which it contains. The diff against it is detectors/context.py and its new test file.

The contextual identifier patterns were tuned against the ai4privacy evaluation corpus rather than against how documents are written, which is the practice benchmarks/evaluate_quality.py's own integrity notice rules out:

NO HARDCODING: Do not write regex or rules targeting specific strings, names, or artifacts found exclusively in this dataset.
BLIND EVALUATION: Do not inspect the validation split to build rules.

Two problems follow from it.

A label was treated as sufficient evidence

Nothing required the captured value to look like an identifier, so an ordinary English word sitting where the identifier belongs was reported as personal data. On main:

"Please provide identification number before Tuesday"
    -> "before"        NATIONAL_ID
"The applicant Jonathan submitted forms."
    -> "Jonathan"      NATIONAL_ID
"We approved a budget of 1500000000000 lire."
    -> "1500000000000" PAYMENT_CARD
"Your school 100200 is nearby."
    -> "100200"        LOCATION

0.15.0's changelog says the digit constraint was added to the NATIONAL_ID contextual fallback, but it reached only two of the patterns; the national-identity one still matched any six-to-fifteen character run, letters included. Every capture now requires at least one digit.

Several triggers were not labels at all

budget of, contribution of, school, office at, singapore, sending your, proof like a, cung cấp ("provide") and número de impresos ("number of printed forms") do not precede an identifier in any document. They precede one in particular rows of one dataset. They are removed, along with the bare id, serial, reference, identifier and applicant triggers that matched with no number marker at all.

The genuine multilingual labels stay — mã số thuế, căn cước, número de identificación, 身份证号, 护照号, paspor, nomor SIM — and the reference-number family still matches when written as a label: reference number, ticket number, serial no.

Verified both directions

Nine prose cases that must not match and ten labelled identifiers across English, Vietnamese and Chinese that must, all as parametrized regression tests:

"Passport No: X1234567 was issued in Rome."   -> X1234567         NATIONAL_ID
"Tax ID: IT12345678901 for the invoice."      -> IT12345678901    TAX_ID
"Ticket number: TK-9928311 is open."          -> TK-9928311       NATIONAL_ID
"Mã số thuế 0101243150 đã đăng ký."           -> 0101243150       TAX_ID
"Credit card number 4111111111111111 declined." -> 4111111111111111 PAYMENT_CARD

Expect the benchmark number to move

The reported precision figure will fall, because some of these matches were scoring as true positives against the very corpus they were derived from. A detector that fires on "the applicant Jonathan" is not more precise in the field; it scores better on one holdout slice. #59 addresses the harness those numbers come from.

263 passed, 1 skipped
coverage 95.46%
ruff, mypy clean

🤖 Generated with Claude Code

https://claude.ai/code/session_012JiX3zWeEC28kmy5KmAXvf