[AdrianAtZyte]

Fixes #1034. Part of #1265.

[codspeed]

Merging this PR will improve performance by ×13

⚡ 2 improved benchmarks
✅ 8 untouched benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_freshness_pattern_findall_long_digit_run	12,942.5 µs	965.3 µs	×13
⚡	test_freshness_pattern_sub_long_digit_run	12,942.9 µs	965.4 µs	×13

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing AdrianAtZyte:freshness-stray-numbers (13daf0a) with master (a3d3b89)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.03%. Comparing base (663348c) to head (13daf0a).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1410   +/-   ##
=======================================
  Coverage   99.03%   99.03%           
=======================================
  Files         239      239           
  Lines        3833     3841    +8     
=======================================
+ Hits         3796     3804    +8     
  Misses         37       37           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[serhii73]

Thanks! This fixes #1034 (03/06/2020a gives 2020-06-03). The catch is that a relative date followed
by any unrelated number is now rejected, and in search_dates that is common prose:

from dateparser.search import search_dates

for s in ["Posted 3 days ago 15 comments", "John Doe 2 hours ago 3 replies",
          "Reviewed 3 weeks ago 12 people found this helpful", "Answered 2 years ago 7 upvotes"]:
    print(s, search_dates(s, languages=["en"]))

On master these give 3 days / 2 hours / 3 weeks / 2 years ago. On this branch (no RELATIVE_BASE)
they give 6 days / ~13 hours / 6 weeks / 4 years ago: the offset is computed from a wrong base. Once 3 day ago 15 is
rejected, search splits the chunk. parse_item parses 3 day in one candidate split and sets
RELATIVE_BASE on parser._settings without restoring it, so the 3 day ago of the next split is
computed from that. The leak is old, but this change is what sends ordinary text through it. With a
RELATIVE_BASE, some chunks are lost instead: "Posted 3 days ago | 15 comments" and
"Last seen 3 days ago at 5" go from 3 days ago to None.

The same rejection hits auto-detected French: search_dates("Il y a 5 min") is translated as
year 1 5 minute, and the stray 1 now makes it January 5th (master: 5 minutes ago). parse() also
loses the day for yesterday at 5, ayer a las 9, вчера в 12 or 3 days ago 10.30 (#1450 recovers
some of these).

Would it work to keep rejecting bare numbers before the quantity (the #1034 shape) and ignore the ones
after the last one? For example, in _are_all_words_units:

matches = list(PATTERN.finditer(date_string))
if matches:
    end = matches[-1].end()
    date_string = date_string[:end] + re.sub(r"\d+", "", date_string[end:])
date_string = PATTERN.sub("", date_string)

With that, the tests in this PR still pass, 03/06/2020a and the 1st of last month are still handled,
and the search_dates and parse() examples above go back to master's results. The
parser._settings leak in search.py parse_item is probably worth fixing on its own too: restoring
the settings after the re-parse also fixes the doubled offsets.

[serhii73]

Code review (claude code-review, at 822ca19). No data-loss or crash found; the full suite passes on this head (24479 passed). The rejection of a unit-less number before a relative phrase is the intended change.

Non-blocking notes:

• The fix is partial and order-dependent: a leading number is now rejected (2020 3 days ago -> None), but digits after the last unit are still dropped silently (3 days ago 15, 5 days ago 2020 and tomorrow 2025 parse as the plain relative date). The new test 3 days ago 15 pins that behaviour, which is the opposite of the PR title.
• When a time is detected, \d+ joins the skip list, so a stray number plus any time token disables the check (2020 3 days ago is None, but 2020 3 days ago 5pm still parses and drops 2020).
• Inputs that parsed before now return None (2020 3 days ago, 12 3 days ago, 3 ago 2 days, 5 yesterday). Please add tests for the rejected cases, since the headline behaviour is currently untested.
• has_time is a new required positional argument on the private _parse_date and _are_all_words_units; a default of False keeps subclasses working.
• The skip list is matched by prefix with re.match, and PATTERN is run twice per candidate on the parse() hot path.